DiffPlantCT: a training-free, annotation-free approach to cross-species plant CT image segmentation Authors: Weizhen Liu, Yihao Fan, Haowei Zhao, Chang Chi, Yihong Wu, Hao Lu, Weijuan Hu, Xiaohui Yuan Published: 2026 ABSTRACT AbstractTraditional deep learning-based plant computed tomography (CT) image segmentation methods require a large amount of high-quality manually labeled data for model training specific to each species, leading to substantial labor costs and poor adaptability to new species. These limitations hinder the application of CT imaging in large-scale cross-species plant phenotyping analysis. Therefore, developing annotation-free and training-free plant CT image segmentation methods is of significant research and application value in reducing research costs and promoting the efficiency of cross-species analysis. To achieve this, we introduce an unsupervised zero-shot segmentation framework for cross-species plant CT images, DiffPlantCT. It is a 2D-to-3D framework that first segments all 2D slices and then assembles them in their original order to generate a 3D CT segmentation. For each slice, this framework directly constructs discriminative clustering features by combining the general semantic priors provided by the self-attention layers in a pre-trained stable diffusion model with the intrinsic grayscale distribution of original image, thereby completely avoiding the need for manual annotations. The method ultimately outputs segmentation results solely through unsupervised clustering, achieving zero-shot generalization without any model training or fine-tuning. To evaluate the feasibility of DiffPlantCT in cross-species segmentation, we benchmark the segmentation performance on two public datasets (walnut fruit and barley spike) and two self-collected datasets (wheat spike and rice panicle). The results show that DiffPlantCT achieved the best performance, with a 41.6% improvement in overall mIoU compared to the state-of-the-art unsupervised method. For the first time, we demonstrate annotation-free, training-free segmentation of cross-species plant CT images successfully.Keywords: Plant CT segmentation, Annotation-free, Training-free, Stable diffusion Introduction Analyzing the internal structural traits of plants is an important step for botanists, crop breeders, and ecologists to understand plant growth patterns [1–4]. In recent years, computed tomography (CT) technology has gained attention for its non-destructive and non-invasive characteristics [1, 5–7], offering a better way to observe the internal structure of plants while preserving the integrity of the sample. Producing high-precision segmentation from CT images [8–11] can enable microscopic visualization and quantitative analysis of plant structures, therefore has significant value for advancing plant biology, crop breeding, and precision agriculture. Traditional deep learning-based segmentation methods for plant CT images predominantly rely on supervised learning paradigms, requiring large amounts of high-quality voxel-level manual annotations for model training [1, 12–14]. This dependency creates key obstacles. First, annotating complex 3D plant structures in CT scans is extremely time-consuming and demands specialized domain expertise. On one hand, this stems from the intrinsic complexity and phenotypic variability of plant morphology (e.g., dynamic structural changes across different growth stages, stringent requirements for fine-grained boundary delineation, and the need for sophisticated 3D spatial cognition of plant organ structures) [11, 15]. On the other hand, 3D CT data is composed of tens to hundreds of stacked 2D slices, and 3D segmentation annotations are usually completed slice by slice [16]. The annotation of a single 3D plant CT image typically takes several hours, which is significantly more time-consuming than 2D image annotation [17]. Second, plants exhibit significant interspecific heterogeneity in tissue texture and morphological structure. Models trained on a single species struggle to adapt to such interspecific structural differences, resulting in poor generalization performance when applied to unseen species [18, 19]. Consequently, researchers are compelled to invest substantial efforts into re-annotating large datasets and retraining models for each new target species. This severely impedes the scalability of cross-species plant phenotypic analysis and hinders the broader application of artificial intelligence in plant science. To alleviate the cross-species annotation burden, researchers have explored various data-efficient learning strategies. Transfer learning is a widely adopted approach that leverages the knowledge acquired by pre-trained models on one task and transfers it to related yet distinct new tasks [20, 21]. For instance, Cassity et al. [21] demonstrated the feasibility of using Mask R-CNN models pre-trained on the COCO dataset as the backbone for onion root CT instance segmentation with minimal species-specific annotations. Other learning paradigms like semi-supervised and few-shot learning offer alternative effective solutions. Semi-supervised learning reduces reliance on extensive species-specific annotations by fusing a small set of high-quality manual labels with large-scale unlabeled plant CT volumetric data, leveraging pseudo-label generation to exploit the informational value of unlabeled data [22]. Few-shot learning enables the model to generalize to unseen species with typically 1–5 labeled samples of the target species; it relies on meta-learning and cross-species morphological prior knowledge to avoid costly large-scale re-annotation for each new species. Successful applications like plant 3D instance point cloud reconstruction [19] and plant pests and diseases segmentation from RGB leaf images [23] have been reported. But these paradigms still require some manual labels, along with subsequent model training or fine-tuning. Thus, developing an annotation-free and training-free framework for cross-species plant CT image segmentation remains an open problem. Unsupervised zero-shot segmentation learning appears to be a highly promising approach [24–27]. Its core strength lies in its ability to learn generalizable and transferable feature representations, allowing direct segmentation of unseen categories without additional training or labeled data (see Fig. 1). In explorations of this direction, diffusion models represented by the pre-trained stable diffusion have demonstrated remarkable performance [27–30]. Beyond segmentation, the rich generative priors embedded in these models have also been successfully harnessed for fully unsupervised image restoration tasks, such as blind denoising, artifact removal, and enhancement of consumer-device photographs, where high-quality ground truth is unavailable [31–33]. These works demonstrate that diffusion priors can effectively separate image structure from noise without paired clean data. Leveraging massive pre-training-derived general visual priors, inherent self-supervised structured learning capabilities of the diffusion process, and precise localization advantages of self-attention mechanisms, these models hold promise for extracting segmentation masks of any objects in arbitrary images without task-specific training [30]. Among them, DiffSeg has exhibited impressive zero-shot segmentation capabilities on 2D RGB images from the COCO and Cityscapes datasets [27]. This method utilizes attention tensors from the unconditioned self-attention layers of stable diffusion models, employing an iterative merging process that measures Kullback–Leibler (KL) divergence between attention maps to form valid segmentation masks. However, our preliminary experiments indicate that direct application of DiffSeg to 2D plant CT slice segmentation yields unsatisfactory results (see Section 3 Results). We attribute this to two key factors. First, unlike RGB images in the COCO and Cityscapes datasets, plant CT slices lack rich color and texture variations, relying on grayscale values to encode structural information with weak grayscale differences between distinct organs and tissues. KL divergence, which captures distribution differences based on distinct statistical variations [34], fails to recognize such subtle changes, leading to failed iterative merging of semantically distinct regions. Additionally, plant CT images suffer from low signal-to-noise ratio (SNR) due to imaging artifacts. KL divergence amplifies noise-induced distribution fluctuations, causing the algorithm to misclassify noise variations as meaningful inter-category differences or overlook real structural distinctions, further reducing spatial grouping accuracy. Comparison between unsupervised zero-shot segmentation and supervised segmentation. “Seen" and “Unseen" refer to categories that the model has encountered before and those it has not encountered previously. The checkmark and cross represent successful and failed outcomes. Supervised segmentation relies on a large amount of manually annotated data for training and struggles to handle categories outside of training. In contrast, unsupervised zero-shot segmentation does not require annotations or training, yet it can still effectively segment images from unseen categories To this end, we propose DiffPlantCT, an annotation-free, training-free 2D-to-3D segmentation framework for cross-species plant CT images. The central concept of DiffPlantCT is to jointly exploit intrinsic image cues, namely, the internal attention structure of pretrained stable diffusion models and the natural grayscale contrast induced by tissue density differences in CT data, to achieve fully unsupervised segmentation. Our primary research objective is to enable accurate 3D semantic partitioning of plant organs and tissues across multiple species without any manual annotation or task-specific training. This work is driven by two key observations. First, we note that the maximum activations in the self-attention tensors of a stable diffusion model can reliably highlight plant foreground structures (organs and tissues) while inherently suppressing background noise, thereby providing a strong prior for slice-wise segmentation. Second, the grayscale variations in CT volumes, which originate from differences in organ and tissue densities, can serve as a natural inter-tissue contrast enhancer that compensates for the absence of color and texture information in CT data. By fusing these two complementary information sources and applying unsupervised clustering, single-slice segmentations are produced and subsequently integrated into a coherent 3D plant model. The overarching research aim is to establish an annotation-free, and training-free 3D segmentation paradigm for plant CT imaging, thus eliminating the dependence on costly labeled datasets and fostering scalable cross-species comparative studies. Materials and methods Plant CT datasets Public datasets The walnut fruit dataset [16] consists of 42 CT scans with a resolution of . Each scan corresponds to a distinct walnut fruit sample, thus containing inherent individual variability. The scans were acquired using a custom-built X-ray CT scanner, the FleX-ray scanner, developed by XRE nv (https://xre.be/) at the Centrum Wiskunde & Informatica (CWI) in Amsterdam, Netherlands. During scanning, detached walnut fruits were fixed by being embedded in foam mounts, and image data were acquired by rotating the sample while keeping the cone-beam microfocus X-ray source and flat panel detector stationary. For each orbit, 1201 projections were captured during the sample’s continuous full rotation, with the first and last projections at the same position resulting in a angular increment. Binned raw projections measured 768 × 972 pixels, with each binned pixel sized 149.6 × 149.6 and a total detection field of view of 114.89 × 145.41 . The source voltage and power were set to 40 kV and 12 W, respectively, adjusted to maximize projection contrast and prevent detector saturation. The barley spike dataset [35] comprises 11 CT scans, each from a distinct barley accession. Scanning was performed using an X-ray nanotomograph SKYSCAN 2214 CMOS EDITION (Bruker Corporation, Heidelberg, Germany). Due to the limited internal space of the scanner, each barley spike was detached from the plant prior to measurement, placed in a plastic holder, and fixed to a disposable plastic platform with adhesive tape to reduce motion artifacts caused by spike rotation during image acquisition. The scanner’s voltage and current were set to 70 and 200 , respectively, with a specimen platform rotation step of 0.9 degrees. Each 2D slice had a pixel size ranging from 75 to 95 , and the 11 3D CT scans yielded a total of 10,433 2D slices. Self-collected datasets The wheat spike and rice panicle datasets are self-collected, with each comprising 10 CT scans and a resolution of . Data acquisition was performed using the self-developed plant X-ray CT system, Shennong-CT V1.0, developed by Institute of Genetics and Developmental Biology, Chinese Academy of Sciences (IGDB, CAS, Beijing, China) [36]. Notably, two key differences exist between the acquisition processes of these self-collected datasets and the other two publicly available ones. Firstly, since the Shennong-CT V1.0 provides sufficient scanning capacity to accommodate an entire potted plant on its scanning platform. Our self-collected datasets were acquired from in vivo wheat spikes and rice panicles without the need for detachment for ex vivo imaging. All data were collected from samples of different varieties and grain filling stages. Secondly, during scanning, the plant remains stationary while the flat-panel detector and X-ray source rotate synchronously. This design reduces motion artifact noise that would otherwise arise from plant rotation. The main scanning parameters of the CT imaging system are detailed in Table 1. Table 1The main CT scanning parameters for self-collected datasetsParameterValueField height per scan135 mmField diameter150 mmProjection image resolution512 × 512Spatial resolution0.3 mmSlice spacing0.3 mmNumber of slices per scan450Single rotation angle0.6°Maximum scanning voltage130 kVThe left side represents the parameters of the acquisition device, while the right side corresponds to their respective values The main CT scanning parameters for self-collected datasets The left side represents the parameters of the acquisition device, while the right side corresponds to their respective values Annotation We conducted slice-wise pixel-level annotation on 3D CT scans from the aforementioned public and self-collected datasets. These datasets cover 42 walnut scans, 11 barley spike scans, 10 wheat spike scans, and 10 rice panicle scans. Figure 2 shows samples from these four datasets along with the corresponding expert annotations. Note that these annotations were used solely for model evaluation. Example images from four plant CT datasets. The top row shows the original images, and the bottom row shows the corresponding labels. The walnut fruit data has been cropped to highlight the internal tissue structure All annotations were performed by three domain experts using ITK-SNAP 4.2.0 [37]. The ITK-SNAP allows users to simultaneously view and manipulate image slices at the same location in the axial, coronal, and sagittal orthogonal planes, providing convenience while ensuring annotation accuracy. In the walnut fruit segmentation task, we defined it as a four-class problem: background, shell, membrane, and kernel. This contrasts with [38], which disregarded the membrane category. Since the membrane structure is closely connected to both the shell and the kernel, ignoring it may lead to incorrect results in subsequent phenotype measurements. For the barley spike, wheat spike, and rice panicle datasets, we frame them as binary classification tasks with the goal of segmenting all grains from the background. Data preprocessing Let be target 3D plant CT scan, where D, H, and W denote the depth, height, and width dimensions, respectively. To adapt CT data to the pre-trained stable diffusion model, we first slice X to obtain all 2D slices. The acquisition protocol may vary depending on the dataset (e.g., different scanning directions). We standardize all inputs by slicing along the axial plane. Formally, we define the set of axial slices S as Eq. (1): Each slice s in S is a single-channel grayscale image, whereas the stable diffusion model is designed for three-channel RGB inputs. To meet this requirement, we convert s into pseudo-RGB format by replicating its single channel across the three color channels, resulting in a transformed set S' as shown in Eq. (2): Finally, these pseudo-RGB images are subjected to intensity normalization. DiffPlantCT DiffPlantCT independently processes each 2D slice image s' ∈ S', and assigns a semantic label c∈ C to each pixel, where C denotes the set of target tissue categories. The predicted 2D segmentation masks are then reassembled into a 3D segmentation volume by preserving their original spatial order along the axial direction. As shown in Fig. 3, DiffPlantCT processes the input image with two branches: maximum attention map extraction and grayscale map sampling. It then performs semantic assignment to generate slice segmentation masks. Finally, the CT segmentation is obtained by performing slice stacking on all the slice segmentation. An overview of the proposed DiffPlantCT. a The overall structure of the framework is presented. DiffPlantCT performs segmentation on all the CT slices. After obtaining the segmentation for all the slices, they are stacked back into the original CT format image. b The DiffPlantCT segmentation process. For each slice, there are two branches: one uses a pre-trained stable diffusion model to obtain the maximum attention map, while the other samples the grayscale map. The components obtained from these two branches are fused, followed by pixel assignment to generate the segmentation mask Maximum attention map extraction. We use a pre-trained stable diffusion model to obtain the maximum attention map, with the time step t set to 300. First, as shown in Fig. 4, after passing through all the transformer layers in stable diffusion model, a total of 16 4D attention tensors A are generated from s' as Eq. (3): where each encodes the attention weights from spatial position (I, J) to other positions. These tensors correspond to four different spatial resolutions: . Architecture diagram of the stable diffusion model. Each module containing a transformer layer generates a 4D attention tensor with a different resolution We perform attention aggregation on all the attention tensors. High-resolution maps yield the most detailed yet fragmented segmentations, while lower-resolution maps provide more coherent segmentations. Therefore, attention aggregation strikes a balance between consistency and detail, which has been demonstrated in DiffSeg [27]. First, we upsample the last two dimensions of each to a unified resolution of 64× 64, yielding . Through Eq. (4), these rescaled tensors are then aggregated with weighting to obtain an aggregated attention tensor : where , . is the fixed aggregation ratio for each attention tensor, calculated as . Finally, we derive the maximum attention map by computing the maximum response across the last two dimensions for each spatial position (I, J) as shown in Eq. (5): Compared to KL divergence in DiffSeg, this operation captures the strongest pixel-wise attention correlation for each location. The self-attention layers in large-scale pretrained diffusion models spontaneously capture structured visual representations without any segmentation supervision [29]. In effect, the various attention heads in these layers function as emergent functional communities that, through their pre-training-driven self-organization, collectively encode coherent object-region patterns [39, 40]. Consequently, foreground regions such as plant tissues exhibit more pronounced and consistent attention patterns than background or noise points, and yield higher values in , as visualized in Fig. 5b. Comparison of attention maps for foreground and background pixels, along with the display of grayscale distributions. a The original image along with two example pixels (red belongs to foreground, green belongs to background). b The attention maps corresponding to the example pixels in a, with the marked points in the image indicating the positions with the strongest response. c The spatial grayscale distribution of the original image. The pixels are mapped to different colors based on the grayscale values Grayscale map sampling. We sample the grayscale map from s\prime. Since differences in plant tissue density are reflected as variations in grayscale values, this variation can enhance the contrast between tissue pixels. Figure 5c shows the grayscale distribution of different tissues in a walnut fruit slice (e.g., shell, membrane, and kernel). We sample the pixels from all positions in s\prime and apply threshold τ to the grayscale values of these sampled pixels to obtain G, as illustrated in Eq. (6): Semantic assignment. From the preceding steps, we obtain two complementary feature representations: and G. These maps are concatenated at the pixel level to form a feature matrix suitable for clustering. First, in Eq. (7), flatten both maps into column vectors: They are then horizontally stacked to construct the feature matrix as Eq. (8): We apply standard K-Means clustering to F to partition pixels into k groups (the number of K-Means iterations N is set to 10), as shown in Eq. (9): where denotes the centroid of the k-th cluster. The label vector is then reshaped to a 64× 64 mask and upsampled to the original resolution, yielding a preliminary segmentation . Due to the random initialization of K-Means, the assignment of cluster indices to plant tissue categories is arbitrary. To enforce consistency across slices, we reassign labels based on the mean grayscale value of each cluster in the original slice s. This allows the cluster labels to be mapped to tissue semantic classes according to their density order. Let be the set of cluster indices in M'. For each , we use Eq. (10) to compute its mean intensity over the corresponding pixels in the original slice s: where is the set of coordinates in M' where the label value is c. represents the number of coordinates in . We then sort by ascending to obtain , as Eq. (11): A mapping function ϕ is defined such that . The final segmentation mask is obtained by applying ϕ element-wise to M', as shown in Eq. (12): Slice stacking. The above pipeline is applied to each pseudo-RGB slice s' ∈ S', producing a sequence of 2D segmentation masks , each of size H× W. As Eq. (13), these masks are reassembled into a 3D volume by stacking them along the depth dimension according to their original spatial order: Evaluation metrics Following prior works [24–27, 41–43], we use mean intersection over union (mIoU) to measure segmentation performance. The background is often meaningless; therefore, to accurately reflect the model’s segmentation performance on the target of interest, the mIoU calculation excludes the background category. Furthermore, considering that grain boundary segmentation affects subsequent downstream tasks, we report 95% Hausdorff Distance () [44] and Average Symmetric Surface Distance (ASSD) [45] on the wheat spike, rice panicle and barley spike datasets, which are two evaluation indicators commonly used to judge the performance of boundary segmentation. Specifically, referring to DiffSeg [27], we emphasize auxiliary prompts (AP), auxiliary images (AX), and unsupervised adaptation (UA). AP refers to the need for prompt input, such as box prompts in MedSAM [26] or text prompts in CLIP [46]. AX signifies that the method requires additional reference images or a pool of synthetic images. UA indicates that the specific method requires unsupervised training. Methods that do not require UA are considered zero-shot [27]. In the overhead experiments, we adpoted three evaluation metrics: inference time, floating-point operations (FLOPs), and GPU memory. Among them, inference time and FLOPs were computed using standard time and PyTorch-OpCounter (thop) [47] library, respectively, while GPU memory usage was obtained via the “nvidia-smi” command. Experimental setup Baselines. We selected six representative unsupervised segmentation methods as baselines for comparison. PiCIE [41] incorporates geometric consistency as an inductive bias to learn invariance and equivariance under photometric and geometric variations. STEGO [42] performs unsupervised semantic segmentation by extracting feature correspondences. MaskCLIP [24] provides a reliable source of supervision for dense prediction tasks, enabling annotation-free segmentation. Both ReCo [25] and DiffSeg [27] exhibit strong zero-shot transfer capabilities: the former introduces a retrieve-and-co-segment strategy, while the latter merges self-attention maps into valid segmentation masks using KL divergence. MedSAM [26], trained on 1,570,263 image–mask pairs, is included for its strong zero-shot segmentation capability. In addition, we also considered three classical image segmentation methods: Otsu [48], Region Growing [49], and Watershed [50]. Implementation details. All experiments were conducted on the Linux CentOS 7 operating system. The deep learning models were implemented based on the PyTorch 2.6 framework [51], with CUDA 12.4, and the programming language version was Python 3.9. In terms of hardware, the experiments were carried out on a server equipped with an Intel Xeon Platinum 8358P CPU (@ 2.60GHz) and a NVIDIA RTX 4090 GPU with 24 GB memory. We use the stable diffusion model V1.4 [28] to obtain self-attention maps for subsequent processes. All clustering algorithms are derived from the official implementations of Sklearn [52]. The methods used in the experiment all reassign each predicted mask to the ground truth mask using the method described above (see section DiffPlantCT). Testing data details. A total of 42 CT scans of walnut fruits were used. To emphasize the segmentation of walnut tissues (shell, membrane, and kernel), 100 consecutive slices were selected from the middle portion of each scan. These slices encompass all walnut tissues, resulting in a total of 4,200 2D slices. 11 CT scans of barley spikes were used, with a total of 10,433 2D slices. For the wheat spike and rice panicle datasets, patching was applied to enhance the stable diffusion model’s focus on the foreground regions, with patches sized . A total of 107 wheat spike patches were used, totaling 13,696 2D slices, while 120 rice panicle patches were used, amounting to 15,360 2D slices. All CT slices were resized to for model inference. After segmenting all the 2D slices, they are reassembled in their original order to form a 3D CT volumes for metric evaluation. Results Comparison with State-of-the-art methods For models that do not require training, we directly performed inference on the test dataset. Text-dependent approaches such as ReCo, we standardized the input prompts across the four datasets to ensure a fair comparison. The prompts used were: “wheat”, “rice”, “walnut shell”, “walnut membrane”, and “walnut kernel”. The K-Means algorithm adopted the k-means++ initialization strategy with a random seed of 42. Table 2 presents the segmentation results of DiffPlantCT and baseline methods on three crop datasets. PiCIE and STEGO rely on unsupervised training. Among them, PiCIE learns features through data augmentation and clustering, but the available augmentation strategies for CT slices are limited, leading to its failure across the three crop datasets. STEGO relies on visual models pre-trained on natural images, such as DINO [53], resulting in limited adaptability to certain crop datasets. For example, it achieves only 27.91% mIoU on the wheat spike dataset. For zero-shot segmentation models, Both MaskCLIP and ReCo require text prompts for guidance. However, due to the absence of text pairs describing CT plant tissues, the vision–language models struggle to align the textual descriptions with the plant structures. Consequently, their overall mIoU scores on the crop datasets are only 37.64% and 13.04%, respectively. Notably, due to domain differences, the general medical segmentation model MedSAM cannot effectively adapt to crop structures, achieving an overall mIoU of only 29.99%. DiffSeg is a zero-shot baseline that does not rely on external semantic guidance, yet it is sensitive to noise and tissue contrast variations in CT images. As a result, the attention maps of the foreground and background become very similar at object boundaries, making it difficult to effectively separate grains. All three classical image segmentation methods require some manual parameter tuning. Among them, Otsu performs the best, achieving relatively good scores on the barley spike dataset, but exhibiting poor performance on other datasets. Region Growing, affected by noise in CT slices, is the worst-performing method among all the baselines. In contrast, DiffPlantCT does not require any additional auxiliary input or model training, yet achieves the best segmentation performance on these datasets: in terms of overall performance, it surpasses the STEGO by 30.14% in overall mIoU; for boundary segmentation, it reduces the overall and ASSD by 8.89 and 2.22, compared with MedSAM. Table 2Comparison with the state-of-the-art methods on wheat spike, rice panicle and barley spike datasetsRequirementsWheat spikeRice panicleBarley spikeOverall MethodAPAXUAmIoU ↑ HD ↓ ASSD ↓ mIoU ↑ HD ↓ ASSD ↓ mIoU ↑ HD ↓ ASSD ↓ mIoU ↑ HD ↓ ASSD ↓ PiCIE [41]✗✗✓12.5137.7812.409.6455.7223.736.2478.8837.559.4657.4624.56STEGO [42]✗✗✓27.916.762.0732.4612.754.0587.1013.101.9449.1610.872.69MaskCLIP [24]✓✗✗33.5627.464.7927.9548.2212.6551.4013.833.5137.6429.846.98ReCo [25]✓✓✗9.3322.936.8210.2731.478.8619.5152.1614.9413.0435.5210.21MedSAM [26]✓✗✗15.4814.593.7921.336.931.8853.178.322.1129.999.952.59Otsu [48]✗✗✗5.1056.6725.9440.524.461.7773.0221.974.2439.5527.7010.65Region Growing [49]✗✗✗0.0371.6534.900.0697.1642.050.02129.5865.940.0499.4647.63Watershed [50]✗✗✗2.1457.7426.0016.976.912.4518.419.936.8812.5124.8611.78DiffSeg [27]✗✗✗19.249.823.0919.5418.164.0634.7924.127.1324.5217.374.76Ours: DiffPlantCT✗✗✗76.541.030.2573.420.850.2887.941.300.5879.301.060.37We employ auxiliary prompts (AP), auxiliary images (AX), and unsupervised adaptation (UA) to demonstrate the requirements of different methods. HD and ASSD are used to measure the segmentation quality of grain boundariesBold indicates the best performance for each metric Comparison with the state-of-the-art methods on wheat spike, rice panicle and barley spike datasets We employ auxiliary prompts (AP), auxiliary images (AX), and unsupervised adaptation (UA) to demonstrate the requirements of different methods. HD and ASSD are used to measure the segmentation quality of grain boundaries Bold indicates the best performance for each metric Table 3 Comparison of computational efficiency of the evaluated methods on crop datasets. The image sizes have been controlled to be the same, and the computational cost is identical across the three crop datasets. Therefore, we performed averaging. DiffPlantCT takes 1.44 s per slice and 184.68 s to process an entire CT volume. This places it at a moderate level among the baseline methods, achieving a balance between segmentation performance and inference time. Regarding FLOPs and GPU memory usage, our framework consumes more resources than most models, and this overhead is almost entirely attributable to the backbone architecture. However, combined with results shown in Table 2, the increased computational cost of DiffPlantCT indeed brings more significant performance gains. Table 3Comparison of Inference time (s), FLOPs (G) and GPU memory (MiB) for baseline methods on crop datasets. Bold indicates the best performance for each metricMethodInference time ↓ FLOPs ↓ GPU memory ↓ Per slicePer 3D volumePiCIE [41]0.2127.140.61518STEGO [42]0.2633.1622.072,062MaskCLIP [24]0.1215.166.361,210ReCo [25]0.5266.65237.871,676MedSAM [26]1.55198.59159.503,564DiffSeg [27]2.47316.48306.5422,696Ours: DiffPlantCT1.44184.68307.0622,692 Comparison of Inference time (s), FLOPs (G) and GPU memory (MiB) for baseline methods on crop datasets. Bold indicates the best performance for each metric The experimental results on the walnut fruit dataset are shown in Table 4, the requirements are consistent with Table 2. In the multi-classification task, classical methods involve substantial manual intervention and were therefore excluded from this experiment to ensure a fair comparison. The complex internal structure of walnuts poses a greater segmentation challenge, especially for the membrane class, which is often tightly connected to both the shell and the kernel. In this task, STEGO and DiffSeg fail to segment walnut tissues correctly. PiCIE, MaskCLIP, and ReCo can identify the correct number of tissue types, but their segmentation quality remains poor, with mIoU scores of 3.12%, 8.70%, and 11.57%, respectively. MedSAM performs relatively well on walnut kernel with an IoU score of 66.19% but shows limited performance on the shell and membrane structures, with IoU scores of 5.85% and 8.84%, respectively. Despite the high difficulty of the task, DiffPlantCT achieves the best performance in segmenting the membrane structures and significantly outperforms other models in the segmentation of both shell and membrane. In general, DiffPlantCT achieves an mIoU of 50.09%, surpassing the best baseline MedSAM by 23.13%, demonstrating a clear performance advantage. Notably, DiffPlantCT performs low-resolution clustering (feature vectors compressed to 64 × 64), which intuitively poses a potential risk of losing fine structural details (e.g., walnut membrane). However, the results in Tab. 4 shows that, despite lower segmentation accuracy on the membrane compared to other tissues, our model still outperforms all baseline methods. This suggests that, while this operation may sacrifice some precision, it overall ensures robust performance. Furthermore, the computational overhead of DiffPlantCT on the walnut dataset also lies at an intermediate level among the baseline methods, achieving a balance between performance and overhead. It is worth noting that, because the model architecture and input size are kept the same, the FLOPs and GPU usage on the walnut dataset produce results consistent with those on crop datasets (Tab. 3). Table 4Comparison with the state-of-the-art methods on walnut fruit datasetRequirementsWalnut fruitInference time ↓ FLOPs ↓ GPU memory ↓ MethodAPAXUAShell ↑ Membrane ↑ Kernel ↑ mIoU ↑ Per slicePer 3D volumePiCIE [41]✗✗✓3.222.833.323.120.24122.350.61518STEGO [42]✗✗✓––––––––MaskCLIP [24]✓✗✗9.124.4312.568.700.1995.346.361,210ReCo [25]✓✓✗12.128.8113.7911.570.57283.62237.871,676MedSAM [26]✓✗✗5.858.8466.1926.964.492,248.94159.503,564DiffSeg [27]✗✗✗––––––––Ours: DiffPlantCT✗✗✗57.0720.8672.3450.091.62813.07307.0622,692Bold indicates the best performance for each metric Comparison with the state-of-the-art methods on walnut fruit dataset Bold indicates the best performance for each metric To provide a more intuitive comparison between DiffPlantCT and other methods, we present the 3D visualization results in Fig. 6. In the wheat spike and rice panicle examples in Fig. 6 top, the segmentation results of the compared methods exhibit severe grain adhesion, particularly for DiffSeg. This issue is less pronounced in the barley example in Fig. 6 middle, as barley spike has larger grain spacing. In contrast, DiffPlantCT performs excellently in crop segmentation, with fewer adhered grains. Overall, the segmentation results of DiffPlantCT for all three crops are closer to the ground truth, making it the best model. Unlike the crop segmentation tasks above, the walnut fruit requires segmentation of its all internal structures. Due to the tight connection between the membranous tissues and other structures, even the human visual system finds it difficult to distinguish them. The results in Fig. 6 bottom show that existing methods struggle to accurately separate these internal structures. In contrast, DiffPlantCT performs remarkably well, effectively distinguishing all types of internal tissue in walnuts. 3D visualizations of the prediction results for different segmentation methods. From top to bottom: wheat spike, rice panicle, barley spike and walnut fruit. The walnut sample is cropped to visualize the segmentation results of the internal structures Ablation studies Performance of different clustering methods. Our approach generates segmentation masks using unsupervised clustering, so it is critical to choose a clustering algorithm suitable for plant CT images. We evaluate the clustering performance of commonly used clustering algorithms (the clustering object is the fusion of the maximum attention map and the grayscale map), including K-Means [54], Gaussian Mixture (GM) [55], Spectral Clustering (SC) [56] and HDBSCAN [57]. We present the results in Fig. 7. It can be observed that K-Means and SC are competitive, performing exceptionally well on the wheat spike, rice panicle, and barley spike datasets, significantly surpassing other clustering algorithms; on the more complex walnut fruit dataset, K-Means outperforms the other algorithms, achieves an absolute 2.3% mIoU improvement over the second-best algorithm, SC. Notably, although different clustering methods exhibit varying performances, overall, they all achieve commendable levels, which indirectly reflects the robustness of the components we have proposed. Overall, K-Means exhibits the best performance, hence we use it as the clustering algorithm in the remaining experiments. Experimental results of different clustering methods on four plant CT datasets. The horizontal axis represents the datasets, and the vertical axis represents the mIoU scores. Different clustering algorithms are distinguished using different color bars Effectiveness of framework components We conduct ablation experiments to evaluate the contribution of each component in DiffPlantCT. For all four plant CT datasets, we test three configurations: clustering with only the maximum attention map, clustering with only the grayscale map, and clustering with the pixel-wise fused combination of both (see Table 5). The baseline in the first row corresponds to DiffSeg, which does not leverage max-pooled attention maps to mitigate noise. First, compared to DiffSeg, using only the maximum attention map improves performance across all datasets, verifying that our max-pooling strategy effectively suppresses noise in self-attention tensors. When using only the grayscale map, we observe substantial gains over the attention-only configuration, especially on the wheat, rice, and barley spike datasets. On these relatively simple structures with clear tissue contrast, the grayscale map already provides sufficient pixel-level intensity cues for accurate segmentation, leading to mIoU values nearly identical to the full fused model. However, the advantage of the fused model becomes evident on the more challenging walnut fruit dataset, where internal structures exhibit weak grayscale contrast and complex spatial layouts. Here, the full model outperforms the grayscale-only variant by an absolute 1.77%. These results demonstrate the complementary roles of the two components: the grayscale map provides robust basic contrast for simple tissues, while the maximum attention map introduces additional structural guidance to resolve ambiguous regions where grayscale intensity fails to distinguish boundaries. The fusion of both thus ensures consistent performance across both simple crop spike structures and complex multi-tissue organs, validating the necessity of the proposed design. Table 5Ablation study on effectiveness of different components of DiffPlantCTK-MeansMaximum attention mapGrayscale mapWheat spikeRice panicleBarley spikeWalnut fruit✗✗✗19.2419.5434.79–✓✓✗32.3929.2649.198.97✓✗✓76.5873.4187.9448.32✓✓✓76.5473.4287.9450.09The values in the table represent the mIoUBold indicates the best performance for each metric Ablation study on effectiveness of different components of DiffPlantCT The values in the table represent the mIoU Bold indicates the best performance for each metric Hyper-parameter studies We investigate the effect of hyper-parameters on segmentation quality. Table 6 shows the settings adapted to the four datasets after experimentation. These values are used for producing the results in Table 2 and Table 4. Here we mainly explore the sensitivity of the time step t of stable diffusion model, grayscale threshold τ , and number of K-Means iterations N. When studying the influence of a certain hyper-parameter, we fix other hyper-parameters as the settings in Table 6. Table 6Hyperparameter configuration of DiffPlantCTNameWheat spikeRice panicleBarley spikeWalnut fruitTime step (t)300300300300Threshold (τ )2800100120–Iterations (N)10101010t represents time steps of the stable diffusion model, τ represents the grayscale threshold of grayscale maps, and N represents the number of iterations of the K-Means clustering algorithm Hyperparameter configuration of DiffPlantCT t represents time steps of the stable diffusion model, τ represents the grayscale threshold of grayscale maps, and N represents the number of iterations of the K-Means clustering algorithm Time step t. We investigated the adaptability of different diffusion time steps in stable diffusion models to plant CT images. Figure 8 presents the corresponding wheat spike heatmap visualizations at different time steps . When t=100, the maximum attention map is of low quality and fails to effectively identify the wheat region. As t=200, the maximum attention map is able to localize the wheat region, but the boundaries remain insufficiently refined. As t increases from 300 to 500, the quality of wheat boundary regions continues to improve. Overall, increasing the time step yields better maximum attention maps, particularly from 100 to 200, where the wheat regions become more salient and noise is reduced. When the time step reaches 300, the maximum attention map is already able to accurately localize the wheat region, along with less noise interference. Moreover, an increase in the number of time steps leads to higher computational cost. When t increases from 100 to 500, the corresponding computation time rises from 0.68 s to 1.04 s. Therefore, considering the computational cost, we set the time step to 300. Visualization of maximum attention maps extracted at different time steps t. The time below each image shows the time spent on the corresponding time step for a single slice Grayscale threshold τ . The CT datasets of wheat spike, rice panicle, and barley spike all contain substantial noise. To enhance the quality of the sampled grayscale images, a grayscale threshold τ is introduced. The addition of the grayscale threshold slightly sacrifices the model’s automatic generalization ability, resulting in an improvement in segmentation performance. Three candidate thresholds are selected for each crop dataset: for wheat spike, for rice panicle, and for barley spike. The experimental results are shown in Fig. 9. On the wheat spike dataset, setting τ to 2200 and 2800 both lead to significant improvements in the model’s segmentation performance, demonstrating the feasibility of applying a grayscale threshold. However, when τ is set to 3400, the mIoU score decreases, indicating that an excessive threshold can degrade model performance. On the rice panicle dataset, setting τ to 90 and 120 improves performance, whereas increasing it to 130 causes a decline. The barley spike dataset shows a comparable trend, where a suitable threshold range of 90–120 enhances segmentation performance, but an excessive threshold of 150 leads to performance deterioration. Therefore, it is reasonable to set the τ of the grayscale maps for wheat spike, rice panicle, and barley spike within the ranges of 2500–3500, 70–100, and 100–150, respectively. Experimental results of grayscale threshold study. a, c show the experimental results on the wheat spike, rice panicle, and barley spike datasets, respectively Number of K-Means iterations N. We select wheat spike and walnut fruit as representative datasets to study the impact of different K-Means clustering iteration numbers N on model performance and computational time. For both datasets, N is set to . On the wheat spike dataset, as N increases from 3, the mIoU of DiffPlantCT continues to improve, indicating that more iterations refine the accuracy of K-Means clustering. However, when N reaches 15, the mIoU stabilizes, suggesting that K-Means has converged. On the walnut fruit dataset, the effect of different iteration numbers is more pronounced. Increasing N from 3 to 5 leads to a significant improvement in K-Means performance. Similarly, when N reaches 10, the segmentation performance of DiffPlantCT remains unchanged. Regarding computational time, both datasets show the same trend: as the number of iterations increases, the time required for K-Means clustering also rises. Therefore, considering both performance and computational cost, we select N = 10 as the optimal iteration number for K-Means clustering in our framework. The experimental results are presented in Fig. 10. Experimental results of K-Means iteration study. The solid blue line represents the mIoU score and the red dotted line represents the clustering time expenditure of K-Means on a single image. a and b Show the experimental results on the wheat spike and walnut fruit datasets, respectively Failure cases We present some failure cases of DiffPlantCT. As shown in Fig. 11, two typical failure modes are observed in slice segmentation. First, when two different plant tissues exhibit similar grayscale distributions, particularly in the regions where they are in contact, boundary ambiguity can occur. In such cases, the model may struggle to accurately delineate tissue boundaries (Fig. 11, first row). Second, over-segmentation may occur when there is substantial background noise surrounding foreground pixels. As shown in the second row of the figure, DiffPlantCT can mistakenly interpret noise as foreground, resulting in spurious segmentation outputs. Future work will further enhance the model’s robustness by tackling these two types of scenarios. Some failure cases of DiffPlantCT. From left to right, each row shows the original image, the prediction result from DiffPlantCT, and the ground truth. The light blue square boxes highlight the failure regions, and local zoom-ins are provided to show the details Application of DiffPlantCT on monitoring rice grain growth To demonstrate the practical utility of our method for plant research, we apply DiffPlantCT to the non-destructive monitoring of rice grain growth dynamics within a panicle, a process otherwise unobservable due to the enclosed structure of the panicle. Longitudinal CT scans of a single rice panicle were acquired at eight time points over 37 days, spanning from the early grain-filling stage to physiological maturity. Each scan was processed with DiffPlantCT to perform grain segmentation. As shown in Fig. 12a, the segmented grain masks are visualized in chronological order, revealing the continuous expansion of grains inside the panicle over time. Non-destructive monitoring of rice grain growth via DiffPlantCT segmentation. a 3D visualization of grain expansion over time. b Richards growth curve of average grain volume. The horizontal axis represents days after grain filling, and the vertical axis represents the corresponding average grain volume of the panicle For the segmentation results at each time point, all valid grains are extracted, and their individual volumes are computed. Specifically, we group voxels belonging to the same grain using 6-connected components, then multiply the number of voxels by the scale factor derived from the scanning configuration (in Table 1) to obtain the volume of each grain. The average grain volume at each time point was then fitted to the Richards equation [58] to generate a quantitative growth curve (Fig. 12b). This curve captures key traits including growth rate, time of maximum growth rate, and final volume, enabling researchers to characterize grain filling dynamics in vivo. Discussion Evaluation methodology design and rationale We design a targeted multi-dimensional evaluation methodology to objectively and comprehensively quantify the segmentation superiority of DiffPlantCT over comparative methods, and to ensure experimental conclusions align with the actual demands of agricultural phenotypic analysis rather than merely pursuing single-metric optimization. All quantitative results supporting this evaluation are presented in Tables 2 and 4, covering both cereal crop spike and walnut fruit tissues to verify cross-species and cross-organ generalization of each method. This evaluation system abides by the norms of mainstream segmentation research [24–27, 41–43], and balances regional overall segmentation accuracy and fine-grained boundary segmentation quality [44, 45], two core performance dimensions, to avoid one-sided evaluation caused by single metric bias. We adopt mean Intersection over Union (mIoU) with the background category excluded as the primary indicator for the core quantitative metric of segmentation performance. As shown in Tab. 2, this metric clearly differentiates method performance. PiCIE [41] achieves an overall mIoU of only 9.46% on three cereal datasets due to destructive data augmentation, while traditional methods (Region Growing [49], Watershed [50]) achieve only 0.04% and 12.51% overall mIoU limited by CT noise and low contrast. Tab. 4 further validates this metric for nut fruit segmentation. Zero-shot methods (MaskCLIP [24], ReCo [25]) only reach 8.70% and 11.57% mIoU on walnut data, whereas DiffPlantCT attains 50.09% mIoU on walnut and 79.30% overall mIoU on cereals, proving its strong cross-species regional segmentation capability. Since boundary segmentation precision directly affects downstream phenotypic tasks, single mIoU cannot reflect boundary fitness, 95th percentile Hausdorff Distance () and Average Symmetric Surface Distance (ASSD) are introduced as dedicated boundary metrics. These metrics quantify boundary deviation between predictions and ground truth. eliminates outlier interference [44], while ASSD reflects the average boundary displacement [45]. Methods with poor boundary performance (STEGO [42], DiffSeg [27]) show extremely high (10.87, 17.37) and ASSD (2.69, 4.76) on cereals, and MedSAM [26] suffers from large boundary errors due to domain mismatch. In contrast, DiffPlantCT reduces and ASSD to 1.06 and 0.37 respectively, achieving precise boundary delineation. Its superior mIoU on walnut tissue components (Table 4) further confirms its fine-grained boundary discrimination ability. Following DiffSeg [27], we classify baseline methods by three practical attributes (marked in Tables 2 and 4): auxiliary prompts (AP), auxiliary images (AX), unsupervised adaptation (UA). Methods relying on AP/AX/UA perform poorly across datasets and are hard to deploy in resource-limited agricultural scenarios. DiffPlantCT avoids these constraints and achieves optimal performance on all metrics, verifying its strong practicality. Overall, this multi-dimensional evaluation system integrates regional accuracy, boundary refinement and practicality analysis, and results in Tables 2 and 4 fully validate DiffPlantCT’s core advantages in cross-species and cross-organ plant CT segmentation. Advantages and applications of DiffPlantCT DiffPlantCT introduces, for the first time, an unsupervised zero-shot inference paradigm for cross-species plant CT image segmentation, overcoming the inherent limitations of conventional deep learning methods that rely on large-scale annotated data and task-specific fine-tuning. Existing supervised approaches depend on extensive annotated datasets for model training and optimization [1, 12–14], making them poorly applicable in plant imaging scenarios where annotations are scarce. In contrast, our framework enables cross-species segmentation without any additional model training or fine-tuning for the target plant CT tasks, providing a universal solution through a lightweight architecture and addressing the gap in zero-shot plant CT segmentation. The term “training-free” in this study requires a clear definition, specifically referring to not modifying model weights for the target plant and not performing any domain-adaptive training [27]. DiffPlantCT solely involves the reasonable integration of general visual priors from a pre-trained foundational model [28] with the intrinsic intensity information of the images. This does not imply inference from randomly initialized models and represents a compliant design within zero-shot visual research. Leveraging its annotation-free and training-free characteristics, DiffPlantCT combines both technical advantages and practical applicability. From a technical perspective, the framework maintains the pre-trained model weights fixed throughout, eliminating the substantial computational cost and manual annotation burden associated with conventional fine-tuning. To address the poor adaptability of DiffSeg [27] to plant CT images, DiffPlantCT employs max pooling on the stable diffusion self-attention tensors in place of the KL divergence to enhance foreground information and suppress noise. The integrated grayscale image further improves contrast between tissue pixels. The performance is supported by both the grayscale features and the prior knowledge embedded in pre-trained models. In terms of engineering application, the framework is well suited for large-scale plant image analysis and real-time deployment scenarios. Benefiting from its annotation-free and fine-tuning-free merits, it is capable of directly processing massive cross-species and multi-scanner CT data, which is further verified by our experimental results (Tables 2 and 4). Concretely, the experimental datasets in this paper cover self-collected data and publicly available datasets, where wheat and rice data are acquired by a single CT scanner, and barley and walnut data [16, 35] are collected by other types of CT scanners. Even under such heterogeneous scanning setups, the model still yields stable segmentation performance and presents basic cross-device generalization behavior, which matches the demands of high-throughput plant phenotyping analysis and large-scale crop breeding screening in modern agriculture [59, 60]. Moreover, the lightweight inference pipeline features low computational consumption and fast inference speed, eliminating the dependence on high-performance servers. It can be embedded into CT scanning equipment or edge hardware to achieve real-time scan-and-segmentation processing, making it applicable to practical scenarios such as laboratory detection and field in-situ analysis. Limitations and future research directions Although DiffPlantCT exhibits significant advantages in training-free inference, cross-species, and cross-scanner generalization, experimental results and the failure cases illustrated in Fig. 11 reveal some core limitations. These shortcomings stem from the inherent trade-offs of the training-free paradigm and are also constrained by current bottlenecks in zero-shot segmentation and 3D vision technologies [26]. At the same time, they help further define the precise boundaries of the terms “training-free”. Firstly, there exists a structural defect in spatial modeling. The framework adopts a slice-by-slice 2D segmentation manner without constructing an explicit 3D spatial correlation mechanism, which disrupts the spatial continuity of CT volumetric data. This may lead to misalignment of segmentation results between adjacent slices during 3D reconstruction and coherence distortion, failing to fully mine the topological structural features of 3D plant tissues [14]. Secondly, the robustness of feature perception is insufficient, with reliance on CT grayscale radiation features. When two plant tissues exhibit highly similar intensity distributions, or the diffusion attention is unstable, the model tends to produce blurred boundaries during segmentation. Furthermore, when complex background noise is mixed into the foreground region, the noise may be incorrectly classified as foreground tissue, leading to over-segmentation. Third, the framework suffers from limited flexibility in engineering deployment. Although the grayscale hyperparameter τ is introduced to boost segmentation performance, as validated in Fig. 9, its dataset-specific manual tuning undermines the feasibility of fully automated deployment across diverse plant species and scanning scenarios. Such manual parameter configuration also compromises segmentation stability when confronted with drastic scanning parameter variations or severe imaging artifacts. Lastly, the proposed model utilizes pre-trained models and CT intensity information for plant structure segmentation. Its deployment requires running Python inference pipelines and standard image preprocessing, creating operational barriers for field plant researchers without computational background and necessitating basic technical training. To address the aforementioned core limitations, future optimization efforts can focus on following directions to further enhance model performance and practical applicability. First, a 3D spatial correlation module [61, 62] can be developed to integrate features from adjacent slices, compensating for the spatial continuity limitations of 2D segmentation and improving 3D tissue reconstruction accuracy. Second, multi-modal feature fusion and attention optimization mechanisms [63, 64] can be considered to enhance the model’s robustness to noise and tissues with similar grayscale values. Third, an adaptive hyperparameter adjustment strategy [65, 66] can be designed to enable data-driven automatic tuning of the grayscale hyperparameter τ , eliminating manual intervention and enhancing the model’s fully automated deployment across diverse scenarios. As for advancing the practical application of this method in botanical and agricultural research, future work may package DiffPlantCT as a graphical user interface (GUI) tool or integrate it as a functional module into large language model (LLM) intelligent agents. Through a streamlined visual interface, researchers can accomplish plant segmentation simply by uploading images, without technical expertise or coding. Beyond the above limitations in algorithm design and practical deployment, this study still has experimental shortcomings in dataset selection. Constrained by the data accessibility, we only selected cereal spikes with simple morphological structures and walnut fruits with complex internal architectures for model validation. The current experiments can only provide a preliminary verification of the method’s effectiveness and are insufficient to fully demonstrate its cross-species generalization ability. Nevertheless, benefiting from the generic semantic representations of pre-trained diffusion models and inherent tissue density differences, DiffPlantCT requires neither species-specific optimization nor additional retraining for new categories. In principle, it possesses the potential to generalize to other fruits, plant roots, and whole-plant segmentation tasks. Further evaluations on broader plant species and organs will be conducted in future work to solidify its scalability. Conclusion In this paper, we verify that plant CT image segmentation can be achieved without training on pixel-level annotations or instance data for target categories. We introduce DiffPlantCT, the first unsupervised zero-shot segmentation framework for cross-species plant CT images. DiffPlantCT generates segmentation results for all CT slices individually, and then assembles them in their original order to achieve full 3D CT image segmentation. For each slice, it fuses the maximum attention map and grayscale map at the pixel level to build discriminative feature objects for unsupervised clustering. In the benchmark tests on four plant CT datasets, DiffPlantCT demonstrated robust segmentation performance, with a 41.6% improvement in overall mIoU compared to the state-of-the-art method. Extensive ablation experiments are also conducted to validate the rationality of each design module. Despite the competitive performance, DiffPlantCT has inherent limitations in spatial modeling, feature perception and manual grayscale threshold tuning. Future work can target these drawbacks to further boost model performance and reduce manual intervention. Acknowledgements We sincerely thank colleagues at the Institute of Genetics and Developmental Biology, Chinese Academy of Sciences, for their assistance with image data collection, and Chang Liu and Guangyu Lan from Wuhan University of Technology for their valuable input on image annotation and figure refinement, respectively. Author contributions WL managed the project, contributed to conceptualization and methodology, supervised the work, acquired funding, and drafted and edited the manuscript. YF conducted the investigation, organized the data, developed the methodology, and drafted the original manuscript. HZ and CC organized data, and participated in the development of the methodology. YW participated in the validation of the methodology. HL reviewed and edited the manuscript. WH and XY supervised the work and provided resources in this study. All the authors read and approved the manuscript. Funding This work was supported by the National Natural Science Foundation of China (Grant Nos. 32570481 and 32200331). Data availability The datasets and implementation code of the DiffPlantCT framework are open-sourced on GitHub at https://github.com/WeizhenLiuBioinform/DiffPlantCT_Zero-Shot_Plant_CT_Segmentation. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Weizhen Liu, Email: liuweizhen@whut.edu.cn. Weijuan Hu, Email: wjhu@genetics.ac.cn. Xiaohui Yuan, Email: yuanxiaohui@whut.edu.cn. References Associated Data Data Availability Statement The datasets and implementation code of the DiffPlantCT framework are open-sourced on GitHub at https://github.com/WeizhenLiuBioinform/DiffPlantCT_Zero-Shot_Plant_CT_Segmentation.