Differentiating sub-centimeter lung metastases in colorectal cancer by deep learning: a multicenter retrospective study Authors: Xinyi Gao, Xiao Ma, Zhiyuan Zhang, Jie Yuan, Qianling Li, Pengwen Zheng, Min Lv, Dening Ma, Jingjing Sun Published: 2026 ABSTRACT BackgroundEarly diagnosis of sub-centimeter lung metastases is critical for timely decision-making and improved prognosis in patients with colorectal cancer. The diagnostic evaluation of indeterminate sub-centimeter lung nodules in colorectal cancer patients remains a crucial challenge. We aim to develop and validate a deep learning model for differentiating sub-centimeter lung metastases noninvasively.MethodsOur retrospective study included 1335 colorectal cancer patients with pretreated sub-centimeter lung metastases from 3 centers and 1335 benign lung nodules from one center. The primary cohort comprised 1194 patients, who were randomly assigned (8:2) to training and internal validation cohorts. Two external validation cohorts (EVC) consisted of 101 (EVC1) and 40 (EVC2) patients. The deep learning framework employed a fully automated VNet-based segmentation model for non-contrast computed tomography (CT) scans, integrated with a ResNet18 classifier for discriminating whether sub-centimeter lung nodules are benign or metastatic. Moreover, stepwise validation of subgroups was performed according to the maximum diameter of the nodules (10, 9, 8, 7, 6, 5, ≤ 4 mm).ResultsThe automatic segmentation model achieved a dice coefficient of 0.825. In primary cohort, the accuracy by radiologists was 0.705. The AUC in deep learning model showed 0.953 (95%CI: 0.937–0.967), 0.906 (95%CI: 0.874–0.926), and 0.951 (95%CI: 0.938–0.965) in internal and two external validation cohorts. Furthermore, the stepwise validation demonstrated that the diagnostic performance remain stable as the nodule maximum diameter shrinks when size≥5 mm.ConclusionThe deep learning model can accurately and non-invasively distinguish between benign and metastatic sub-centimeter lung nodules in patients with colorectal cancer.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12880-026-02508-8. Background Colorectal cancer (CRC) remains the third most commonly diagnosed cancer and the second leading cause of cancer related mortality worldwide [1]. Among people initially diagnosed with CRC, 20% will present with distant metastases, resulting in a poor prognosis [2]. Lung is one of the most prevalent sites of metastasis, second only to the liver, manifesting in approximately 15% of patients with distant metastasis [3, 4]. For the patients diagnosed with lung metastases (LM), only about 9–12% of patients are eligible for local radical intent treatments, including surgical resection, stereotactic body radiation therapy (SBRT) and ablation therapy [5]. The size of LM at initial diagnosis will influence treatment decisions, and is a crucial factor for both local control and survival [6, 7]. Early and accurate identification of small LM is essential for quick intervention and individualized treatment, thereby further improving the survival outcomes of patients. Chest CT scans are recommended for diagnosis and management of CRC patients with LM at baseline and follow-up period [8]. CT images can provide more information in detecting sub-centimeter LM than PET-CT and X-ray [9]. However, due to small diameter and similar morphological features, it is difficult to differentiate sub-centimeter LM from benign nodules on CT. Approximately 16% of indeterminate lung nodules are proved to be metastatic during follow up [10], and some of these nodules are sub-centimeter in size. Given that sub-centimeter LM are likely to be missed or over-diagnosed by CT [11, 12], the interpretation of such lung nodules represents a considerable challenge. Deep learning (DL) is an end-to-end approach that automatically extracts high-throughput information from specific images, significantly improving the accuracy and robustness of diagnostic models [13, 14]. Compared to classical machine learning algorithms, DL is better suited to tackle complex issues such as large-scale image classification, where models scale exponentially as the quantity and dimensionality of data grows [15, 16]. To the best of our knowledge, no studies have used DL models to distinguish sub-centimeter LM in CRC patients. This study aimed to evaluate whether deep learning analysis can extract relevant features from non-contrast chest CT scans to improve diagnostic accuracy of sub-centimeter LM. We launched a multicenter collaboration for the development and validation of a deep learning model to differentiate sub-centimeter LM in CRC patients, followed by stepwise validation of model stability according to the nodule maximum diameter. Methods Patients This retrospective multicenter study was conducted in accordance with the Declaration of Helsinki. The study was approved by the local ethics review board of all the contributing centers, with written informed consent waived (IRB-2022-149, IRB-2111246-26, IRB-2019-750-105-01 respectively). The study population enrollment and workflow are shown in Fig. 1. The inclusion criteria for CRC patients with sub-centimeter LM were: (1) histopathologically confirmed CRC(age>18years); (2) LM diagnosed by histopathology or imaging evidence (increase in size) of adequate follow-up(ranging from 3 to over 6 months) to ensure definitive clinical confirmation [10]; (3) at least one LM detected by helical CT≤10 mm in maximum diameter when LM was first diagnosed; (4) no evidence of additional primary cancers. An experienced radiologist (X, 10 years work experience) re-examined both the initial and follow-up chest CT scans to confirm the diagnosis of LM. The exclusion criteria were as follows:1) history of other primary cancers; 2) lack of follow-up CT images to confirm LM; 3) severe CT image artifacts precluding reliable analysis. A total of 1335 CRC patients were retrospectively collected from three medical centers (a primary cohort and two independent external validation cohorts), with one definitive sub-centimeter LM randomly selected per patient for analysis. The primary cohort included 1194 consecutive patients from center A from August 2006 to July 2021, which were randomly split into training and internal validation cohorts (IVC) at a 8:2 ratio. The two independent external validation (EVC)cohorts included 101 patients from center B from January 2010 to August 2022 and 40 patients from center C between January 2018 and March 2022. Fig. 1Flowchart of the enrolment population Flowchart of the enrolment population For the benign cohort, a total of 1335 benign nodules were randomly selected from 1335 anonymous patients without a history of cancer at primary center A between October 2015 and October 2018. The inclusion criteria required nodules to have a maximum diameter of ≤ 10 mm, a regular shape, and well-defined margins. Conversely, nodules presenting with any suspicious malignant features including but not limited to irregular shape, spiculation, lobulation, pleural indentation, or cavitation were strictly excluded. To ensure diagnostic ground truth, all selected nodules underwent a standard surveillance program comprising annual chest CT scans, and their benign nature was confirmed by stable size and morphology during a minimum of three years of follow-up imaging. The total 2670 nodules, including 1335 LM and 1335 benign nodules were assigned to primary training cohort, IVC, EVC 1, and EVC 2 in a 1:1 ratio. In order to further explore the generalizability of DL model on the sub-centimeter nodules of different sizes, we carried out a stepwise validation on the subgroups according to the maximum diameter of the nodules. The nodules in the internal and two EVC groups were divided into seven subgroups (maximum diameter was 10, 9, 8, 7, 6, 5, and ≤ 4 mm). Baseline clinical report review To establish a realistic clinical baseline for comparison, we retrospectively evaluated the diagnostic accuracy of the initial routine clinical reports from the primary cohort. These chronologically generated reports were originally signed off by more than 30 certified radiologists (> 5 years of independent practice) who were entirely blinded to the final histopathological or clinical outcomes. For diagnostic performance assessment, a correct diagnosis was defined as an explicit, definitive radiological interpretation of pulmonary metastasis. Conversely, reports omitting metastasis, recommending routine follow-up without a definitive impression, or misclassifying the nodule as a benign lesion or primary lung malignancy were classified as incorrect. Imaging protocol All patients underwent standard unenhanced chest CT during the initial staging or the follow-ups. The detailed CT imaging protocols are presented in Supplementary. Image preprocessing The lung parenchyma segmentation model developed by Johannes Hofmanninger, which is based on the U-net architecture [17], was utilized in this study. All the images were normalized into [0, 1] and resampled into 1 mm×1 mm size. Automatic ROI segmentation V-Net has been demonstrated to segment large-scale medical images effectively, so we have constructed a convolutional neural network utilizing the V-Net architecture to facilitate the automatic segmentation of regions of interest (ROI) (Fig. 2). The original model’s convolution kernel size was reduced from 5 × 5 × 5 to 3 × 3 × 3 to minimize parameters, and batch normalization was added to each convolution layer for faster training. The initial input data size was set to 1 × 128 × 128 × 32, after applying two convolution kernels with 16 channels, the size changed to 16 × 128 × 128 × 32. A ResBlock module with downsampling was then incorporated, featuring three 3 × 3 × 3 convolution cores with 64 channels, and an output size of 64 × 32 × 32 × 8. Then, the third and fourth ResBlock modules with downsampling were added, leading to a final output of 256 × 8 × 8 × 2 before upsampling, which utilized a 3 × 3 × 3 deconvolution kernel. The concat hop connection mechanism was employed to restore lost features during downsampling. Post-upsampling, the output returned to 1 × 128 × 128 × 32. The model underwent pre-training on the LIDC-IDRI dataset, containing 1018 labeled CT scans, with the best-performing model selected after 150 iterations. An experienced radiologist (Y, 15 years work experience) manually delineated ROIs on CT images of 110 lesions, dividing them into 88 for training and 22 for validation. Using the Adam optimization algorithm with a learning rate of 0.001 and a batch size of 16 over 50 iterations, the model achieved a dice coefficient of 0.825, demonstrating effective automatic ROI segmentation. Fig. 2Workflow of the deep learning analysis. This figure shows the workflow of establishing a predictive model for sub-centimeter lung nodules using deep learning, including automated nodule segmentation and a 3D ResNet18-based deep learning modeling process Workflow of the deep learning analysis. This figure shows the workflow of establishing a predictive model for sub-centimeter lung nodules using deep learning, including automated nodule segmentation and a 3D ResNet18-based deep learning modeling process Deep learning analysis The three-dimensional ResNet18 was used to train the deep learning model (Fig. 2 and Figure S1) [18]. The network was adapted for single-channel 3D CT input by modifying the first convolutional layer, and an adaptive global average pooling layer followed by a fully connected layer was used for final classification. Batch normalization was applied after convolutional layers to accelerate training and improve convergence. A patch-based strategy was used for model input: according to the nodule mask information, each lung nodule was cropped as a 28 mm × 28 mm × 28 mm 3D CT patch centered on the lesion, thereby increasing the relative display proportion of the nodule. Transfer learning was applied using the MedMNIST standardized medical image dataset of lung nodules, which included 1849 benign and malignant lung nodules. The model was optimized using the Adam optimizer, and cross-entropy loss was used during training. Weight initialization followed the method proposed by Glorot et al. [19]. The batch size was set to 32, and the initial learning rate was set to 0.001, with all training data iterated for 100 epochs. For subgroup validation, the trained 3D ResNet18 model was directly applied to the validation cases stratified by nodule size, without additional retraining or fine-tuning. According to the measured maximum diameter, nodules in the validation cohorts were divided into seven subgroups: 10, 9, 8, 7, 6, 5, and ≤4 mm. The same preprocessing, ROI cropping, and model inference procedures as the main validation analysis were used for all subgroups. For each subgroup, the model generated the predicted probability of lung metastasis for each nodule. AUC was calculated using the continuous predicted probabilities, whereas accuracy, sensitivity, specificity, PPV, NPV, and F1-score were calculated using a fixed probability threshold of 0.5; nodules with predicted probabilities greater than 0.5 were classified as metastatic nodules. Statistical analysis Continuous variables were described as mean ± standard deviation and categorical variables were described as frequencies and percentages. Normal or non-normal distribution of continuous variables was determined by Kolmogorov-Smirnov test. Levene’s test was used to test for continuous variance alignment. The area under the receiver operating characteristic curve (AUC) with 95% confidential interval (CI) and accuracy were calculated to evaluate the diagnostic performance. Chi-square test and Mann-Whitney test were used to compare the categorical variables, and independent samples t-test was used to compare the continuous variables. Decision curve analysis was used to evaluate the clinical net benefit of the model. Statistical analysis was carried out using SPSS 25 (IBM, Armonk, NY, USA). P-value < 0.05 indicated the statistical difference. Results Baseline characteristics of patients The characteristics of the CRC patients with LM are shown in Table 1. Additionally, the fundamental characteristics of the benign nodule cohort are summarized in Supplementary Table S1. For training cohorts, the mean diameter was 6.84 ± 1.81 mm (benign lung nodules were 4.68 ± 1.53 mm). For IVC group, the mean diameter was 6.85 ± 1.79 mm (benign lung nodules were 4.60 ± 1.37 mm). For EVC1, the mean diameter was 6.52 ± 1.84 mm (benign lung nodules were 4.58 ± 1.51 mm). For EVC2, the mean diameter was 6.93 ± 1.70 mm (benign lung nodules were 4.70 ± 1.31 mm). A highly significant difference in the maximum long diameter was observed between benign nodules and pulmonary metastases across all four cohorts (P < 0.001).In the primary cohort of 1,194 patients, 842 cases were correctly diagnosed during the initial assessment, while 352 cases were not accurately diagnosed, yielding a radiologist’s diagnostic accuracy rate of 0.705. To assess potential variability, diagnostic accuracy was analyzed across three tiers of clinical experience: 5–10 years (0.722), 10–15 years (0.706), and > 15 years (0.702). A chi-square test revealed no statistically significant difference among these groups (χ²=0.219, P = 0.896), demonstrating consistent baseline performance across the clinical department.Furthermore, diagnostic accuracy was evaluated within each nodule size subgroup. A gradual decline in diagnostic performance was observed as nodule size decreased, dropping to a low of 0.609 in the ≤ 4 mm subgroup (Supplementary Table S2). Table 1Characteristics in the primary cohort and two external validation cohortsPTC †(n = 954)IVC ‡(n = 240)EVC1§(n = 101)EVC2¶(n = 40)Age (years)58.64 ± 10.9659.68 ± 10.8154.08 ± 11.9963.03 ± 12.56Primary tumor location Rectum5771436520 Colon356922620 Unknown21510Stage I20900 II8618177 III251695316 IV5381281217 Unknown5916190Gender Male5951596427 Female359813713Maximum diameter (mm)6.84 ± 1.816.85 ± 1.796.52 ± 1.846.93 ± 1.70†PTC= primary training cohort, ‡IVC= internal validation cohort, §EVC1 = external validation cohort 1, ¶EVC2 = external validation cohort 2 Table 1Characteristics in the primary cohort and two external validation cohortsPTC †(n = 954)IVC ‡(n = 240)EVC1§(n = 101)EVC2¶(n = 40)Age (years)58.64 ± 10.9659.68 ± 10.8154.08 ± 11.9963.03 ± 12.56Primary tumor location Rectum5771436520 Colon356922620 Unknown21510Stage I20900 II8618177 III251695316 IV5381281217 Unknown5916190Gender Male5951596427 Female359813713Maximum diameter (mm)6.84 ± 1.816.85 ± 1.796.52 ± 1.846.93 ± 1.70†PTC= primary training cohort, ‡IVC= internal validation cohort, §EVC1 = external validation cohort 1, ¶EVC2 = external validation cohort 2 Characteristics in the primary cohort and two external validation cohorts †PTC= primary training cohort, ‡IVC= internal validation cohort, §EVC1 = external validation cohort 1, ¶EVC2 = external validation cohort 2 Diagnostic validation of deep learning models As shown in Fig. 3; Table 2, the DL model demonstrated robust diagnostic performance for discriminating between benign and metastatic lung nodules, yielding an AUC of 0.992 (95%CI:0.989–0.994) in the primary training cohort(PTC), 0.953 (95%CI:0.937–0.967) in the IVC cohort, 0.906 (95%CI:0.874–0.926) in the ECV1 cohort, and 0.951 (95%CI:0.938–0.965) in the ECV2 cohort.Radar maps depicting the balance in performance of the model, using the eight diagnostic indexes as the dimensions, are shown in Fig. 4A. Fig. 3Diagnostic validation of deep learning models. Figure 3 illustrates the receiver operating characteristic curve of the deep learning models in primary training cohort, internal validation cohort, external validation cohort 1 and external validation cohort 2 Diagnostic validation of deep learning models. Figure 3 illustrates the receiver operating characteristic curve of the deep learning models in primary training cohort, internal validation cohort, external validation cohort 1 and external validation cohort 2 Table 2Validation in different cohorts by deep learningAUC†AccuracySensitivitySpecificityPrecisionF1PPV‡NPV§IVC¶0.9530.8920.8960.8880.8880.8920.8880.895EVC 1*0.9060.8520.8820.8260.8120.8360.8120.891EVC 2**0.9510.8750.9250.8250.8410.8850.8410.917†AUC=area under curve, ‡PPV=positive predictive value, §NPV=negative predictive value, ¶IVC=internal validation cohort, *EVC 1 = external validation cohort 1, **EVC 2 = external validation cohort 2 Table 2Validation in different cohorts by deep learningAUC†AccuracySensitivitySpecificityPrecisionF1PPV‡NPV§IVC¶0.9530.8920.8960.8880.8880.8920.8880.895EVC 1*0.9060.8520.8820.8260.8120.8360.8120.891EVC 2**0.9510.8750.9250.8250.8410.8850.8410.917†AUC=area under curve, ‡PPV=positive predictive value, §NPV=negative predictive value, ¶IVC=internal validation cohort, *EVC 1 = external validation cohort 1, **EVC 2 = external validation cohort 2 Validation in different cohorts by deep learning †AUC=area under curve, ‡PPV=positive predictive value, §NPV=negative predictive value, ¶IVC=internal validation cohort, *EVC 1 = external validation cohort 1, **EVC 2 = external validation cohort 2 Fig. 4A-B Radar charts of the models. These radar charts depicted the balance in performance of the models, using the eight diagnostic indexes as the dimensions (AUC, accuracy, sensitivity, specificity, precision, F1, PPV, and NPV). A demonstrated robust performance across all three validation cohorts (internal validation, external validation 1, and external validation 2). B maintained stable performance (all eight dimensions) across different sub-centimeter size stratified validation cohorts. AUC=area under curve, F1 = F1 score, PPV=positive predictive value, NPV=negative predictive value A-B Radar charts of the models. These radar charts depicted the balance in performance of the models, using the eight diagnostic indexes as the dimensions (AUC, accuracy, sensitivity, specificity, precision, F1, PPV, and NPV). A demonstrated robust performance across all three validation cohorts (internal validation, external validation 1, and external validation 2). B maintained stable performance (all eight dimensions) across different sub-centimeter size stratified validation cohorts. AUC=area under curve, F1 = F1 score, PPV=positive predictive value, NPV=negative predictive value In order to assess the clinical benefit value, we employed decision curve analysis to determine the model score interval that could benefit patients based on the model’s recommendations. According to the decision curve analysis showed in Fig. 5, the three validation cohorts exhibited promising clinical utility. Fig. 5Decision curve analysis of deep learning models. Figure 5 illustrates the decision curve analysis of the deep learning models in internal validation cohort, external validation cohort 1 and external validation cohort 2 Decision curve analysis of deep learning models. Figure 5 illustrates the decision curve analysis of the deep learning models in internal validation cohort, external validation cohort 1 and external validation cohort 2 To interpret the model’s decision-making process, visualization heatmaps were analyzed (Supplementary Figure S2). The feature maps demonstrated that the deep learning model applied varying degrees of attention to different regions of interest; specifically, it focused distinctively on both the internal nodule characteristics and the surrounding microenvironment to differentiate nodule malignancy. Stepwise validation in different sub-centimeter size subgroups We measured the maximum diameter of the nodule three times and took its average value to minimize the measurement error. All the LM in three validation cohorts were divided into the above mentioned seven subgroups, the proportion of each group were 10.5%, 10.0%, 11.0%, 20.5%, 20.2%, 18.6%, and 9.2%, respectively with the maximum diameter gradually decreased. The AUC remains stable in deep learning (Table 3). As the maximum diameter of LMs decreased from 10 mm down to 5 mm, the AUCs remained stable. Although the AUC dropped to 0.807 for nodules ≤ 4 mm, the model still demonstrated higher diagnostic performance, with its lowest accuracy (0.757) remaining substantially higher than that of radiologists in the same subgroup (0.609).Radar maps depicting the balance in performance of the model, using the eight diagnostic indexes as the dimensions, are shown in Fig. 4B. Table 3Stepwise validation in different sub-centimeter size subgroups by deep learningAUC†AccuracySensitivitySpecificityPrecisionF1PPV‡NPV§10 mm0.9740.9000.8750.9250.9210.8970.9210.8819 mm0.9710.8850.9230.8460.8570.8890.8570.9178 mm0.9770.9641.0000.9290.9330.9660.9331.0007 mm0.9800.9210.9610.8820.8900.9240.8900.9576 mm0.9390.8910.9360.8460.8590.8960.8590.9305 mm0.9000.8240.7710.8750.8570.8120.8570.797≤ 4 mm0.8070.7570.6000.9140.8750.7120.8750.696†AUC=area under curve, ‡PPV=positive predictive value, §NPV=negative predictive value Table 3Stepwise validation in different sub-centimeter size subgroups by deep learningAUC†AccuracySensitivitySpecificityPrecisionF1PPV‡NPV§10 mm0.9740.9000.8750.9250.9210.8970.9210.8819 mm0.9710.8850.9230.8460.8570.8890.8570.9178 mm0.9770.9641.0000.9290.9330.9660.9331.0007 mm0.9800.9210.9610.8820.8900.9240.8900.9576 mm0.9390.8910.9360.8460.8590.8960.8590.9305 mm0.9000.8240.7710.8750.8570.8120.8570.797≤ 4 mm0.8070.7570.6000.9140.8750.7120.8750.696†AUC=area under curve, ‡PPV=positive predictive value, §NPV=negative predictive value Stepwise validation in different sub-centimeter size subgroups by deep learning †AUC=area under curve, ‡PPV=positive predictive value, §NPV=negative predictive value Discussion An early and noninvasive diagnosis of sub-centimeter LM plays a crucial role in formulating personalized treatment strategies and improving prognosis. However, there are currently no noninvasive tools capable of accurately predicting the nature of lung nodules in CRC patients. This study demonstrated that DL approaches exhibited good performance in differentiating benign lung nodules from LM, achieving AUC of 0.992, 0.953, 0.906, 0.951 in the PTC, IVC, EVC1 and EVC2, respectively. The DL model also showed superior diagnostic accuracy compared to radiologists (0.852 [lowest] vs. 0.705), demonstrating its potential as a clinical useful adjunct for early detection of sub-centimeter LM before progression occurs. In contrast to previous studies, this research focuses on the more clinically challenging task of identifying sub-centimeter LM. Our approach combines transfer learning and deep learning for automated nodule segmentation, extracts high-dimensional features from a large-scale cohort, and performs size-stratified validation across multicenter datasets to ensure generalizability. In CRC patients, the diagnostic evaluation of sub-centimeter LM remains a crucial clinical challenge. Although indeterminate lung nodules can be diagnosed through histopathological biopsy, including thoracoscopy and needle biopsy, the technical feasibility decreases substantially with diminishing nodule size. For sub-centimeter nodules, biopsy procedures may yield inadequate tissue samples, resulting in reduced diagnostic accuracy, with a concomitant risk of potential complications, such as alveolar hemorrhage and pneumothorax [7]. Some previous studies have reported the limited diagnostic value of PET-CT/MR for sub-centimeter nodules, given that these nodules demonstrate inadequate FDG avidity for definitive assessment [20, 21]. CT scan is the primary modality for detecting and characterizing LM in CRC patients. However, in a cohort study of 521 patients, CT demonstrated limited diagnostic accuracy (61.7%) for detecting LM [22]. Particularly in sub-centimeter lung nodules, the absence of characteristic radiological signs frequently poses diagnostic challenges [23]. For these nodules, a 3–6 months follow-up is typically recommended to monitor changes and clarify their nature, which may increase patient costs and potentially delay therapeutic intervention. In this study, 352 out of 1,194 sub-centimeter LM (29.5%) in the primary center were only diagnosed upon disease progression, having been overlooked during first presentation. Sub-centimeter LM and benign lung nodules are difficult to be distinguished by conventional imaging features. Nevertheless, CT harbors numerous subtle heterogeneous features beyond visual perception, and mining these features may overcome current clinical limitations. Some studies have employed radiomics approaches to extract additional discriminative features for the identification of LM in CRC patients. In a single-center cohort of 194 CRC patients, Hu et al. developed a combined clinical-radiomic nomogram that showed excellent discrimination capacity, with respective AUC values of 0.929 in training and 0.922 in validation sets [24]. A bi-institutional study demonstrated that the CT-based radiomics model achieved consistent predictive performance, with C-index values of 0.793 and 0.730 in training and validation sets [25]. However, the nodule size in the above studies varied from 5 to 20 mm in diameter. For clinical challenging indeterminate sub-centimeter lung nodules, the diagnostic accuracy and reliability of current artificial intelligence (AI) models remain uncertain. Our study has several strengths. First, compared with the previous literature focused on lager nodules, our work uniquely concentrates on sub-centimeter LM (≤ 10 mm), which present greater diagnostic difficulties and clinical management challenges. To our knowledge, this is the first study to differentiate sub-centimeter LM using deep learning approach. Second, we established a robust deep learning model using large-scale and multicenter datasets, comprising 1335 CRC patients, significantly enhancing its generalizability and clinical applicability. Third, we developed an efficient V-Net based network for automated ROI segmentation, reducing reliance on time-consuming manual delineation while maintaining high reproducibility. Another key innovation lies in our in-depth analysis of the model’s performance across sub-centimeter lung nodules of varying diameters. With the LM diameters decreasing from 10, 9, 8, 7, 6, and 5 mm, the AUCs remain stable. Previous studies had shown that both radiologist assessments and radiomics-based diagnoses were affected by nodule size, with diagnostic accuracy declining for sub-centimeter nodules [26, 27]. This aligns with clinical observations, while larger nodules are typically diagnosed correctly, sub-centimeter nodules often pose diagnostic challenges due to their indistinct features. Although existing models demonstrated diagnostic capability for mixed-size nodules, their performance tends to degrade when applied to indeterminate sub-centimeter lesions. In contrast, our model maintained consistent performance across sub-centimeter sizes, achieving an AUC of 0.807 even for the smallest nodules (≤ 4 mm in diameter).This diagnostic performance limitation for nodules ≤ 4 mm is primarily driven by the inherent scarcity of distinct morphological features in sub-5 mm lesions, coupled with technical constraints such as partial volume effects and image noise that obscure critical fine details. To better understand the biological basis of our model’s predictions, we evaluated regional feature attention using heatmaps (Supplementary Figure S2). This interpretability is qualitatively supported by a representative clinical case of a colorectal cancer patient presenting with an initially indeterminate pulmonary nodule (Fig. 6). In this instance, the high-attention zones highlighted by the baseline heatmap—predominantly capturing internal and peripheral structural irregularities—showed a notable spatial overlap with the regions that subsequently exhibited more pronounced growth during a 3-month follow-up. Although a single retrospective observations cannot be definitive, this alignment suggests that the deep learning features may reflect inherent morphological markers associated with biological progression. By offering these localized visual insights, the model holds potential as an auxiliary tool to complement clinical judgment, particularly when managing borderline or radiologically ambiguous lesions. Fig. 6Clinical case demonstration of deep learning-assisted pulmonary nodule evaluation. Baseline CT scan of an elderly female patient with stage IV rectal cancer exhibits an ambiguous 0.9 cm pulmonary nodule with a model-predicted malignancy probability of 0.994. The corresponding baseline heatmap overlay reveals model attention (warm hues) concentrated within the nodule’s internal and peripheral regions, corresponding to subtle structural irregularities. Follow-up CT imaging at 3 months demonstrates nodule enlargement to 1.4 cm, clinically and radiologically confirmed as a metastasis, where the morphological expansion appears more noticeable in the medial aspect Clinical case demonstration of deep learning-assisted pulmonary nodule evaluation. Baseline CT scan of an elderly female patient with stage IV rectal cancer exhibits an ambiguous 0.9 cm pulmonary nodule with a model-predicted malignancy probability of 0.994. The corresponding baseline heatmap overlay reveals model attention (warm hues) concentrated within the nodule’s internal and peripheral regions, corresponding to subtle structural irregularities. Follow-up CT imaging at 3 months demonstrates nodule enlargement to 1.4 cm, clinically and radiologically confirmed as a metastasis, where the morphological expansion appears more noticeable in the medial aspect Nevertheless, the current study has several limitations. First, as a retrospective study, this research may be subject to selection bias. Second, not all sub-centimeter LM had pathological confirmation, particularly benign nodules in the control group, due to the invasive nature of biopsy procedures. Conclusions In conclusion, this study provided alternative noninvasive diagnostic tools based on routine non-contrast chest CT for determining sub-centimeter LM, demonstrating improved accuracy in comparison to radiologists’ assessments. It has the potential to guide optimal management strategies for indeterminate lung nodules detected by initial or follow-up chest CT scans in CRC patients, which contributing to improved individualized treatment outcomes and survival rates in this population. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1