A deep learning framework for lesion-level treatment response prediction in hodgkin lymphoma using PET/CT tensor radiomics Authors: Mahdie Jajroudi, Hossein Jamalirad, Milad Enferadi, Vahid Roshanravan, Farshad Emami, Parham Geramifar, Saeid Eslami Published: 2026 ABSTRACT AbstractBackgroundAccurate prediction of treatment response in Hodgkin lymphoma (HL) is crucial for personalized therapy. The Tensor Radiomics (TR) paradigm advances traditional radiomics by producing and analyzing diverse feature variations, employing tensors computed across multiple parameter combinations to optimize radiomics feature extraction and improve tumor characterization. However, the complexity and variability of these features pose significant challenges for analysis and clinical application in HL research. This study introduces the Tensor Radiomics Network (TR-NET), a deep learning framework that automates feature selection from tensor radiomics data and predicts treatment response. It also incorporates explainable AI techniques to identify the most influential features.MethodsWe analyzed 420 lesions from 70 HL patients using 18F-FDG PET/CT. Tensor radiomic feature sets were generated by varying bin widths for PET (0.2, 0.4, 0.6, 0.8) and CT (10, 25, 50) images. Extracted features encompassed first-order, shape, texture, and filter classes (Laplacian of Gaussian, Wavelet, Square, Square Root, Logarithm, Exponential, Gradient, Local Binary Pattern (LBP) 2D, LBP 3D for each lesion, totaling 13,790 features. We utilized machine learning models, including Random Forest (RF), XGBoost, Support Vector Machine (SVM), and the deep learning model TR-NET for analysis. SHAP (SHapley Additive exPlanations), an explainable AI method, was applied to interpret the predictions of the best-performing model.ResultsTR-NET achieved a predictive performance, with an AUC-ROC of 0.8470 (95% CI: 0.7669–0.9185), a sensitivity of 82.2%, and a specificity of 78.2%, offering a balanced classification characteristic. In comparison, XGBoost achieved an AUC of 0.8160 (95% CI: 0.7457-0.8599), characterized by high sensitivity (93.3%) but lower specificity (66.2%), indicating a tendency towards false positives. Random Forest and SVM also showed high sensitivity (92% and 91.1%), but both suffered from significantly reduced specificity (29.5% and 51.3%).ConclusionTR-NET demonstrates enhanced predictive accuracy and clinical relevance compared to classic machine learning, supporting personalized treatment strategies and the integration of explainable AI in oncology.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40658-026-00838-8.Keywords: Tensor radiomics, Hodgkin lymphoma, PET/CT, Deep learning, Treatment response prediction Introduction Hodgkin lymphoma (HL) is a common malignancy in young adults, with treatment failure or relapse occurring in approximately 10–15% of early-stage and 15–30% of advanced-stage cases, despite advances in therapy [1, 2]. Early prediction of treatment response is crucial for optimizing outcomes, as it enables timely adjustments to therapy, reduces overtreatment in responders, and facilitates escalation in non-responders. This approach supports the growing adoption of risk-adapted treatment strategies in clinical practice and highlights the need for reliable, interpretable biomarkers to guide early response assessment and personalize care in HL [3–5]. 18F-FDG PET/CT plays a key role in HL staging, monitoring, and residual disease evaluation. While Standardized Uptake Values (SUVs) derived from positron emission tomography/ computed tomography (PET/CT) are useful, they can be influenced by various factors, including patient preparation, scan timing, and reconstruction methods [6–9]. Radiomics features (RFs) quantify tumor heterogeneity and offer prognostic information beyond visual assessment [10–12], and in HL, they have been applied to staging, treatment response prediction, and survival modeling [13–22]; however, the robustness and standardization of RFs remain significant challenges due to their stability being influenced by factors such as preprocessing methods, discretization bin widths, segmentation approaches, and variability in scanner hardware or imaging protocols. These factors may lead to the extraction of features that are non-robust to noise or demonstrate inconsistencies, thereby undermining reproducibility and interpretability across studies. The diversity of feature selection methods in radiomics research reflects the challenge of identifying predictive, non-correlated features, as individual features may lack added value or obscure biological meaning [23, 24]. Tensor radiomics (TR) extends radiomics by integrating multidimensionality and multimodality, generating multiple feature values (“flavours”) under varying parameters like discretization binning, modality ratios, and pre-processing filters (e.g., Laplacian of Gaussian, Wavelets) to capture a comprehensive spectrum of tumor heterogeneity across imaging modalities [24, 25]. This is primarily due to the vast number of features it generates, which require extensive feature engineering—a process that is both time-consuming and demands significant expertise [24–27]. Combining TR with deep learning can automate feature selection and improve performance, but interpretability remains a challenge. Integrating explainable AI (XAI) can address this issue, providing insights into how specific tensor radiomics features contribute to predictions, thus improving clinical trust and utility [28, 29]. Among various XAI techniques, SHapley Additive exPlanations (SHAP) is a widely used method that quantifies the contribution of each feature to a model’s prediction. By assigning individual SHAP values to features, this approach helps identify the most influential radiomics biomarkers, offering greater transparency and aiding clinical decision-making. This study evaluates TR’s potential for predicting early treatment response in HL using baseline PET/CT. TR-NET (Tensor Radiomics-NET), a deep learning framework, was developed and applied to analyze TR features derived from various bin widths and filters. By comparing TR-NET’s performance with machine learning and employing SHAP for interpretability, we aim to identify key metabolic and structural biomarkers influencing treatment outcomes, ultimately advancing personalized HL treatment. Materials and methods This study employed a three-methodological approach encompassing radiomics, predictive modeling, and XAI (Fig. 1). To ensure reproducibility and transparency, the study adhered to established reporting guidelines, including the Standards for Reporting of Diagnostic Accuracy Studies (STARD) and the Checklist for Artificial Intelligence in Medical Imaging (CLAIM). Overview of the proposed framework, which consists of three key components: the radiomics pipeline, the modeling pipeline, and the explainable AI pipeline. The radiomics pipeline involves pre-processing PET and CT images to extract tensor radiomics features. The modeling pipeline integrates both machine learning methods (SVM, RF, and XGBoost) and the deep learning-based TR-NET architecture. Tensor radiomics features are used by all these models, with SMOTE applied to address the class imbalance in the training data. The explainable AI pipeline leverages SHAP to interpret the best performance model predictions, identifying the most influential features for treatment response classification. This comprehensive framework combines advanced radiomics, robust modeling techniques, and interpretability to enhance predictive performance and clinical utility Tensor radiomics pipeline Study design and data collection This retrospective, two-center study was conducted at Razavi Hospital (Mashhad, Iran) and Khatam Hospital (Tehran, Iran). Ethical approval was obtained from the Ethical Committee of Mashhad University of Medical Sciences (IR.MUMS.REC.1401.347), and informed consent was obtained from all participants. From 2017 to 2022, approximately 2677 lymphoma patients underwent whole-body 18F-FDG PET/CT scans. Inclusion criteria were: adult HL patients (≥ 16 years old) treated with doxorubicin hydrochloride (Adriamycin), bleomycin sulfate, vinblastine sulfate, and dacarbazine (ABVD) chemotherapy with both baseline (bPET/CT) and post-treatment PET/CT scans. Lesions treatment response was assessed using the Lugano criteria based on post-treatment PET/CT scans [30]. Patient selection is detailed in Fig. 2. PET/CT scans were acquired using a Siemens Biograph 6 (Center I) and a Siemens Biograph mCT (Center II). Follow-up scans were performed on the same scanner as the baseline scan. Patients fasted for at least 6 h pre-injection, maintaining blood glucose below 1.8 g/L. 18 F-FDG was administered at 3.5–4.5 MBq/kg (maximum 450 MBq), with scans initiated approximately 60 min post-injection. PET images were reconstructed using Ordered Subset Expectation Maximization (OSEM) with Point Spread Function (PSF) correction, without TOF, applied consistently across both centers using vendor-specific software for the Siemens Biograph 6 and Siemens Biograph mCT scanner. Digital Imaging and Communications in Medicine (DICOM) image characteristics are detailed in Table 1. Patient selection flowchart. This flowchart outlines the inclusion and exclusion criteria for patient selection in our study. A total of 2,677 lymphoma patients were initially considered. The final cohort included Hodgkin lymphoma patients who underwent both baseline and post-treatment PET/CT scans Table 1Scanner specifications for center I and center IIScanner specificationsCenter ICenter II‌PET matrix size168 × 168200 × 200PET voxel size (column, row, slice thickness) (MM)4.07 × 4.07 × 34.07 × 4.07 × 3CT matrix size512 × 512512 × 512CT voxel size (column, row, slice thickness) (mm)0.97 × 0.97 × 30.97 × 0.97 × 1 Scanner specifications for center I and center II Image segmentation Two experienced nuclear medicine physicians independently segmented lesions on baseline PET/CT scans, manually using 3D Slicer 5.4, with a third expert resolving any discrepancies. Inter-observer agreement was then evaluated using intraclass correlation coefficients (ICC) to ensure reliable texture calculations [6, 31]; lesions were required to have a minimum volume of 64 voxels (4 × 4 × 4), corresponding to 4 cm³ with a voxel size of 4 mm. Lesions smaller than this threshold were excluded. Tensor radiomics feature extraction and harmonization PET voxel values were converted to body weight SUV using standard methods prior to feature extraction. CT and PET images from both centers were resampled to a uniform voxel size to PET images from Center I, and then all images were converted to Nearly Raw Raster Data (Nrrd) format. In this study, we considered two types of tensor radiomics flavours: (1) Discretization flavours, defined by varying bin widths for PET and CT images, and (2) Filter (pre-processing) flavours, derived by applying image transformations such as wavelet and Laplacian of Gaussian (LoG) filtering. Lesions radiomics features were extracted using PyRadiomics 3.1, adhering to image biomarker standardization initiative (IBSI) guidelines from baseline PET/CT. These included first-order statistics, shape features, texture features (grey level co-occurrence matrix (GLCM), neighboring grey tone difference matrix (NGLDM), grey level run length matrix (GLRLM), grey level size zone matrix (GLZLM), grey level dependence matrix (GLDM)), and filter-based features (Laplacian of Gaussian Filter (LoG), Wavelet, Square, Square Root, Logarithm, Exponential, Gradient, Local binary patterns (LBP) 2D, LBP 3D). ComBat harmonization was used to mitigate inter-center variability and ensure consistent feature distributions by adjusting feature means and variances [17, 31]. ComBat offers advantages over methods like z-score normalization, quantile normalization, and Generative Adversarial Networks (GAN)-based approaches by effectively accounting for confounding variables while preserving underlying data structure, making it suitable for complex datasets. Its use of mixed-effects models and empirical Bayes estimation addresses both fixed and random effects in multi-center data. Empirical evidence demonstrates ComBat’s efficacy in reducing inter-center variability, achieving homogeneity without compromising data integrity, thus enhancing downstream analysis robustness [31–33]. Lesions were classified as responding (0) or non-responding (1) based on post-treatment PET/CT scans, using the Lugano criteria’s 5-Point Deauville Score system. Responding (0) included Complete Metabolic Response (CMR) with scores of 1–3 in nodal or extranodal sites, without a residual mass. Non-responding (1) encompassed Partial Metabolic Response (PMR) with scores of 4–5 and reduced FDG uptake from baseline with residual mass(es), Stable Disease with scores of 4–5 and no significant change in FDG uptake, or Progressive Disease with scores of 4–5 and increased FDG uptake intensity from baseline, without additional progression criteria beyond Deauville scores. Extracted features are detailed in Table 2. Table 2Main and sub-categories of radiomics featuresMain categoriesSub-categoriesFirst- OrderMeanMedianStandard DeviationVarianceSkewnessMaximumMinimumKurtosisEntropyRangeEnergyRoot Mean Square (RMS)10th, 90th PercentilesTotal EnergyInterquartile Range (IQR)UniformityShapeElongationFlatnessLeast Axis lengthMajor Axis LengthMaximum 2DDiameter ColumnMaximum 2D Diameter RowMaximum 2D Diameter SliceMaximum 3DDiameterMesh VolumeMinor AxisLengthSphericitySurface AreaSurface VolumeRatioVoxel VolumeGLCM1AutocorrelationCluster ProminenceCluster ShadeCluster TendencyContrastCorrelationDifference AverageDifferenceEntropyDifference Variance,Id6Idm7Idmn8Idn9lmc110lmc2Inverse VarianceJoint AverageJoint EnergyJoint EntropyMCC11Maximum ProbabilitySum AverageSum EntropySum SquaresGLRLM2Gray Level Non-Uniformity,Gray Level Non-Uniformity NormalizedGray LevelVarianceEmphasisLong Run EmphasisHigh Gray Level Run LongRun High Gray Level Emphasis,Long Run Low Gray Level Emphasis,Low Gray Level Run Emphasis,Run Entropy,Run Length Non-Uniformity,Run Length Non-Uniformity Normalized,Run Percentage,Run Variance,Short RunEmphasis,Short Run High Gray LevelEmphasisShort Run Low Gray Level EmphasisGLSZM3Gray Level Non-UniformityGray Level Non-Uniformity NormalizedGray Level VarianceHigh Gray Level Zone EmphasisLarge Area EmphasisLarge Area High Gray Level EmphasisLarge Area Low Gray Level EmphasisLow Gray Level Zone EmphasisSize Zone Non-UniformitySize Zone Non-Uniformity NormalizedSmall Area EmphasisSmall Area High Gray Level EmphasisSmall Area Low Gray Level EmphasisSmall Area Low Gray Level EmphasisZone EntropyZone%Zone VarianceGLDM4Dependence EntropyDependence Non-UniformityDependence Non-Uniformity NormalizedDependence VarianceGray Level Non-UniformityGray Level VarianceHigh Gray Level EmphasisLarge Dependence EmphasisLarge Dependence High Gray Level EmphasisLarge Dependence Low Gray Level EmphasisLow Gray Level EmphasisSmall Dependence EmphasisSmall Dependence High Gray Level EmphasisSmall Dependence Low Gray Level EmphasisNGTDM5BusynessCoarsenessComplexityContrastStrengthFilter Classeswavelet (12LLL, 13LLH, 14LHL, 15LHH, 16HHH, 17HLH, 18HHL, 19HLL)Laplacian of Gaussian (LoG, sigma 0.5,1,2)SquareSquare RootLogarithmExponentialGradient20LBP (2D)LBP(3D)1GLCM = grey level co-occurrence matrix, 2GLRLM = grey level run length matrix, 3GLSZM = grey level size zone matrix, 4GLDM = grey level dependence matrix, 5NGTDM = neighboring grey tone difference matrix, 6Id = inverse difference, 7Idn = inverse difference normalized, 8Idm = inverse difference moment, 9Idmn = inverse difference moment normalized, 10Imc = informational measure of correlation, 11MCC =matthews correlation coefficient, 12LLL=low-low-low, 13LLH=low-low-high, 14LHL=low-high-low, 15LHH = = low-high-high, 16HHH=high-high-high, 17HLH=high-low-high, 18HHL=high-high-low, 19HLL=high-low-low, 20LBP=Local Binary Pattern. Each first and second-order feature was extracted from the original imaging and then applied in all filter classes Main and sub-categories of radiomics features Mean Median Standard Deviation Variance Skewness Maximum Minimum Kurtosis Entropy Range Energy Root Mean Square (RMS) 10th, 90th Percentiles Total Energy Interquartile Range (IQR) Uniformity Elongation Flatness Least Axis length Major Axis Length Maximum 2D Diameter Column Maximum 2D Diameter Row Maximum 2D Diameter Slice Maximum 3D Diameter Mesh Volume Minor Axis Length Sphericity Surface Area Surface Volume Ratio Voxel Volume Autocorrelation Cluster Prominence Cluster Shade Cluster Tendency Contrast Correlation Difference Average Difference Entropy Difference Variance, Id6 Idm7 Idmn8 Idn9 lmc110 lmc2 Inverse Variance Joint Average Joint Energy Joint Entropy MCC11 Maximum Probability Sum Average Sum Entropy Sum Squares Gray Level Non-Uniformity, Gray Level Non- Uniformity Normalized Gray Level Variance Emphasis Long Run Emphasis High Gray Level Run Long Run High Gray Level Emphasis, Long Run Low Gray Level Emphasis, Low Gray Level Run Emphasis, Run Entropy, Run Length Non-Uniformity, Run Length Non-Uniformity Normalized, Run Percentage, Run Variance, Short Run Emphasis, Short Run High Gray Level Emphasis Short Run Low Gray Level Emphasis Gray Level Non-Uniformity Gray Level Non-Uniformity Normalized Gray Level Variance High Gray Level Zone Emphasis Large Area Emphasis Large Area High Gray Level Emphasis Large Area Low Gray Level Emphasis Low Gray Level Zone Emphasis Size Zone Non-Uniformity Size Zone Non-Uniformity Normalized Small Area Emphasis Small Area High Gray Level Emphasis Small Area Low Gray Level Emphasis Small Area Low Gray Level Emphasis Zone Entropy Zone% Zone Variance Dependence Entropy Dependence Non-Uniformity Dependence Non-Uniformity Normalized Dependence Variance Gray Level Non-Uniformity Gray Level Variance High Gray Level Emphasis Large Dependence Emphasis Large Dependence High Gray Level Emphasis Large Dependence Low Gray Level Emphasis Low Gray Level Emphasis Small Dependence Emphasis Small Dependence High Gray Level Emphasis Small Dependence Low Gray Level Emphasis Busyness Coarseness Complexity Contrast wavelet (12LLL, 13LLH, 14LHL, 15LHH, 16HHH, 17HLH, 18HHL, 19HLL) Laplacian of Gaussian (LoG, sigma 0.5,1,2) Square Square Root Logarithm Exponential Gradient 20LBP (2D) LBP(3D) 1GLCM = grey level co-occurrence matrix, 2GLRLM = grey level run length matrix, 3GLSZM = grey level size zone matrix, 4GLDM = grey level dependence matrix, 5NGTDM = neighboring grey tone difference matrix, 6Id = inverse difference, 7Idn = inverse difference normalized, 8Idm = inverse difference moment, 9Idmn = inverse difference moment normalized, 10Imc = informational measure of correlation, 11MCC =matthews correlation coefficient, 12LLL=low-low-low, 13LLH=low-low-high, 14LHL=low-high-low, 15LHH = = low-high-high, 16HHH=high-high-high, 17HLH=high-low-high, 18HHL=high-high-low, 19HLL=high-low-low, 20LBP=Local Binary Pattern. Each first and second-order feature was extracted from the original imaging and then applied in all filter classes To assess the impact of discretization, we selected PET bin widths of 0.2, 0.4, 0.6, and 0.8, and CT bin widths of 10, 25, and 50, based on the Freedman-Diaconis rule as defined in Eq. 1, expert knowledge, and the radiomics literature [24, 34–37]. The Freedman-Diaconis rule calculates optimal bin widths to balance data granularity and noise reduction, using the interquartile range (IQR) of voxel intensities and the number of voxels ( n ). For PET images, with an SUVmean IQR of 4.3–6.68 (median: 5.56) and voxel counts ranging from 64 to 18,566 (median: 197), the rule yielded bin widths of 0.2,0.6, and 0.8, as detailed in Supplementary Table S1. For CT images, with an IQR of 81–170 (median: ~100) and the same voxel range, bin widths ranged from 10,25, and 50. Three sigma values (0.5, 1, and 2) were applied to the LoG filter for TR features [21]. Where h, IQR, and n are bin width, Interquartile Range, and number of voxels. Visualization of treatment response pattern t-Distributed Stochastic Neighbor Embedding (t-SNE) was used to visualize the high-dimensional data in a two-dimensional space. t-SNE is a dimensionality reduction technique that preserves local structure by clustering similar data points while maximizing inter-cluster distances. We used t-SNE to visualize the clustering of responding and non-responding patients. Effective bin widths for PET and CT should ideally result in distinct separation of these two groups, indicating their potential for accurate treatment classification. Modeling pipeline This part included Resampling of Imbalanced Data, modeling, and evaluation. For treatment response assessment modeling, we employed machine learning models (Random Forest (RF), XGBoost (eXtreme Gradient Boosting), Support Vector Machine (SVM)), and the deep learning model (TR-NET). Resampling imbalanced Of the 420 lesions analyzed, 30% (126 lesions) were randomly reserved as an unseen test set, with the remaining 70% (294 lesions) utilized for training and validation; ten-fold cross-validation was performed within this 70% to tune, optimize, and evaluate the models before final testing. The training set was subjected to the Synthetic Minority Over-sampling Technique (SMOTE) to address the class imbalance, which can bias models toward the majority class, hindering generalization and minority class sensitivity. Unlike random over-sampling (which duplicates instances), SMOTE generates synthetic samples by interpolating between existing minority class samples and their nearest neighbors. This maintains minority class diversity while mitigating overfitting associated with duplicating data. Figure 1 illustrates the SMOTE process in a 2D feature space, showing the original imbalanced dataset and the synthetic samples (blue points) augmenting the minority class (orange points) to create a more balanced distribution. Machine learning methods Feature preprocessing was performed to optimize model efficiency and performance. This included removing low-variance features and highly correlated variables. Univariate analysis (t-tests for normally distributed features and Mann-Whitney U tests for non-normally distributed features) identified features significantly associated with treatment response (p < 0.05), which were retained for subsequent modeling. This selection process ensured the inclusion of only the most relevant features [19]. The following machine learning models were employed: RF: A bagging-based ensemble method constructing multiple decision trees from bootstrap samples. RF aggregates predictions to improve accuracy and reduce overfitting. It is effective with high-dimensional data and provides feature importance rankings [38]. XGBoost: A gradient boosting algorithm that sequentially trains decision trees, optimizing residual errors and incorporating regularization to prevent overfitting. XGBoost is known for its efficiency and strong performance on structured data [38]. SVM: A supervised learning method that identifies the optimal hyperplane to separate classes, performing well in high-dimensional spaces and with complex decision boundaries [39]. Model stability and generalizability were ensured through ten-fold cross-validation. Hyperparameter tuning was performed using a grid search within the cross-validation framework to identify optimal model configurations. Deep learning method TR-NET is a two-stage deep neural network (DNN) designed for robust prediction using multi-parametric TR features. In the first stage, modality- and bin-specific subnetworks (referred to as “legs”) independently process the high-dimensional TR feature vectors. Each leg projects its input into a unified, lower-dimensional latent space that captures key radiomics patterns across all flavours, including variations in intensity distribution and spatial heterogeneity that are unique to each imaging parameter and binning scale. The second stage of TR-NET—the “body”—aggregates the outputs from all subnetworks into a shared representation, which is then refined through fully connected layers for final classification. A sigmoid-activated output unit provides a probabilistic estimate of treatment response. The architecture utilizes Scaled Exponential Linear Unit (SELU) activations, dropout regularization, and Mean Squared Error (MSE) loss, which have demonstrated robust convergence and performance in high-dimensional radiomics settings [24, 40]. Figure 3 illustrates the specific implementation of TR-NET in this study, adapted for treatment response prediction in HL. The network comprises seven parallel subnetworks (“legs”), each corresponding to a specific radiomics flavour: four derived from PET images using SUV bin widths of 0.2, 0.4, 0.6, and 0.8, and three from CT images with bin widths of 10, 25, and 50. Each leg receives an input vector of 1,970 TR features, which is processed through two dense layers (128 and 64 neurons) using SELU activation and a dropout rate of 0.5. TR-NET Architecture: A two-stage deep feed-forward neural network that integrates and optimizes radiomics features extracted from PET and CT images. The architecture comprises multiple subnetworks, referred to as "legs," each processing tensor radiomics features corresponding to a specific "flavour" defined by unique bin sizes. For PET images, the flavours correspond to bin sizes of 0.2, 0.4, 0.6, and 0.8, while for CT images, the flavours encompass bin sizes of 10, 25, and 50. The architecture effectively captures the distinct spatial and textural patterns associated with varying intensity discretization levels by dedicating a separate leg to each flavour. The outputs from all "legs" are then concatenated and passed into the "body," which consists of fully connected layers that aggregate these features into a unified representation for final prediction The outputs from all seven legs are concatenated to form a 448-dimensional latent vector (7 flavours × 64 features), capturing complementary patterns across modalities and scales. This fused representation is passed into the second stage—the “body” of the model—consisting of three fully connected layers (448 → 128 → 64 → 1). The final sigmoid-activated node produces a probability score (ranging from 0 to 1) representing the likelihood of treatment response for each lesion. To enhance generalizability and prevent overfitting, the model employs MSE loss, L2 regularization (coefficient = 0.001), and a progressively decreasing dropout schedule. This architecture enables end-to-end learning from high-dimensional TR features while preserving both modality-specific detail and cross-flavour consistency. Ten-fold cross-validation, splitting the 70% training/validation set into 90% training and 10% validation per fold, selected the optimal architecture based on the lowest validation loss and highest accuracy. The final model was retrained on the complete training set before testing. This two-step process ensures robust learning and reliable performance with high-dimensional tensor radiomics features. Models’ evaluation The performance evaluation of our models utilized state-of-the-art metrics, grounded in contemporary literature, to ensure a robust and reliable assessment. The effectiveness of the classifiers was quantified using key performance indicators, including Precision, Sensitivity (Recall), Specificity, Accuracy, and the F1-score. Furthermore, the Area Under the Receiver Operating Characteristic Curve (AUC-ROC) was utilized to provide a comprehensive summary of the model’s ability to differentiate between classes by plotting the true positive rate against the false positive rate across various thresholds. To complement these metrics, Training Time (measured in minutes), defined as the duration of training each model on extracted TR feature vectors using ten-fold cross-validation, was recorded to evaluate computational efficiency, excluding feature extraction and inference times. This evaluation approach effectively balances predictive performance with computational efficiency, providing a thorough understanding of the models’ capabilities. To statistically assess differences in model performance, we applied the DeLong test, a nonparametric method for comparing correlated ROC curves, to evaluate the significance of AUC differences between TR-NET and other classifiers. A p-value < 0.05 was considered statistically significant. Additionally, we reported 95% confidence intervals (CI) for AUC, sensitivity, and specificity to quantify uncertainty and improve interpretability. Explainable AI: the Shapley additive explanations (SHAP) XAI enhances the transparency and interpretability of machine learning and deep learning models by clarifying the post-training prediction generation process, fostering trust, and ensuring compliance with practical standards. Among XAI techniques, SHAP (SHapley Additive exPlanations) excels due to its robust framework, leveraging Shapley values from cooperative game theory to interpret predictions. SHAP models feature attribution as a game, where features are “players,” and the model’s prediction is the “result,” assigning each feature a SHAP value that reflects its average contribution across all possible combinations. This method maintains three key properties: local accuracy (SHAP values sum to the difference between the prediction and a baseline), missingness (uninfluential features receive a zero value), and consistency (SHAP values adjust to reflect model behavior changes). These characteristics emphasize feature importance, rendering SHAP effective for interpreting models, thus enhancing trust and aiding model refinement. Deep SHAP extends this approach to deep neural networks like TR-NET, integrating SHAP values with DeepLIFT (Deep Learning Important FeaTures) via a layered approximation technique, efficiently computing feature contributions through backpropagation, linear combination rules, gradient-based estimates for non-linear components, and reference value propagation across network layers. In this study, SHAP was applied exclusively to interpret the predictions of the highest-performing model, as determined by evaluation metrics, concentrating interpretability efforts on the most robust and clinically relevant framework to reveal the influence of tensor radiomics features on predictive outcomes [41, 42]. In this study, SHAP was applied exclusively to interpret the predictions of the highest-performing model, as identified by evaluation metrics, focusing interpretability efforts on the most robust and clinically relevant framework to uncover the influence of tensor radiomics features on predictive outcomes. All code was implemented in Python using PyTorch 2.5, scikit-learn 1.5.2, and Deep SHAP on the CPU. Results A total of 420 lesions from 70 HL patients were analyzed, with 65.2% (274 lesions) classified as responding and 34.8% (146 lesions) as non-responding according to the Lugano Criteria assessment in post-treatment PET/CT scans. The cohort consisted of 37 males (52.9%) and 33 females (47.1%), with a mean age of 33 years (SD: 16.43) and a mean weight of 76.85 kg (SD: 25.81). At the patient level, 70% (49 patients) responded to treatment, and 30% (21 patients) were non-responsive. Disease stages were distributed as follows: 48.5% (34 patients) Stage II, 17.1% (12 patients) Stage III, and 34.3% (24 patients) Stage IV. Lesion characteristics included a median SUVmax of 13.53 (IQR: 10.33–15.89), a median SUVmean of 6.06 (IQR: 4.3–6.68), and a median metabolic tumor volume (MTV) of 9.87 (IQR: 4.13–15.9). Patient and lesion characteristics are detailed in Table 3. Inter-rater reliability for lesion segmentation and treatment response assessment, measured by ICC between two independent physicians, was 0.87, indicating strong agreement. Table 3Patient and lesions characteristicsPatients’ characteristicsN = 70Percentage (%)Age (mean/std)Weight (mean/std)33/16.4376.85/25.81––Gender(M\F)Male(N)Female (N)373352.85%47.14%ResponseYes: (N)No: (N)592170%30%StageI(N)II(N)III(N)IV(N)0341224048.5%17.14%34.28%Lesions’ characteristicsN = 420SUVmax (median/ Q1, Q3)13.53 (10.33,15.89)SUVmean (median/ Q1, Q3)6.06(4.3,6.68)MTV (median/ Q1, Q3)9.87 (4.13,15.9)Maximum of Hounsfield unit (median/ Q1, Q3)100 (81,170)Number of Voxel197 (110, 423) Patient and lesions characteristics Age (mean/std) Weight (mean/std) 33/16.43 76.85/25.81 – – Gender(M\F) Male(N) Female (N) 37 33 52.85% 47.14% Response Yes: (N) No: (N) 59 21 70% 30% Stage I(N) II(N) III(N) IV(N) 0 34 12 24 0 48.5% 17.14% 34.28% Comprehensive analysis of tensor radiomics features for PET and CT For each lesion, 1,970 radiomics features were extracted per bin width, yielding a total of 13,790 features across all widths. To visualize and explore the distribution of tensor radiomics features in a reduced-dimensional space, we applied t-SNE. This nonlinear technique projects high-dimensional data into two dimensions while preserving local structures. Figure 4 presents the t-SNE representation of radiomics feature distribution for each lesion as responding or non-responding, with colors indicating their respective treatment outcomes. Visualization of Tensor Radiomics Features Using t-SNE.A t-SNE plot illustrates the distribution of tensor radiomics features extracted from PET and CT images for treatment response classification. Each point represents a lesion, with yellow and purple indicating non-responding and responding lesions, respectively. Features corresponding to different bin sizes for PET (0.2, 0.4, 0.6, and 0.8) and CT (10, 25, and 50) were projected into a two-dimensional space to explore clustering patterns across bin widths in the responding and non-responding groups. Although some overlap between groups is observed, partial clustering highlights the potential for differentiating metabolic and structural characteristics related to treatment response The visualization reveals patterns of feature clustering influenced by the chosen bin widths for PET (0.2, 0.4, 0.6, and 0.8) and CT (10, 25, and 50). While lesions from both response categories appear interspersed, distinct regions of higher local density emerge, suggesting underlying differences in metabolic and structural characteristics. Notably, features derived from higher bin widths (PET SUV 0.8 and CT 50) exhibit more pronounced separation, indicating that these discretization settings may enhance response-related feature discrimination. Performance analysis of models Table 4 presents the performance metrics of the predictive models evaluated using tensor radiomics features for testing the lesion dataset, including TR-NET, XGBoost, RF, and SVM. TR-NET achieved an AUC-ROC of 0.8470 (95% CI: 0.7669–0.9185), sensitivity of 82.22% (95% CI: 70.83–91.84), specificity of 78.21% (95% CI: 68.83–87.01), and a training time of 3 min and 43 s. XGBoost obtained an AUC-ROC of 0.8160 (95% CI: 0.7457–0.8599), with a sensitivity of 93.33% (95% CI:0.8478-1 ) and a specificity of 66.23% (95% CI: 55.07–76.40). Random Forest achieved an AUC-ROC of 0.8238 (95% CI: 0.7394–0.8962), with a sensitivity of with a sensitivity of 92% (95% CI: 0.8536-1) and a specificity of 29.49% (95% CI: 0.1899-0.4). SVM reported an AUC-ROC of 0.6947 (95% CI: 0.6400–0.7800), with a sensitivity of 91.11% (95% CI: 82.05–98.00) and a specificity of 51.28% (95% CI: 39.72–62.03). DeLong test p-values comparing TR-NET to other models were 0.53 for RF, 0.4072 for XGBoost, and < 0.001 for SVM. Table 4Performance metrics of predictive models for hodgkin lymphoma treatment response prediction on an unseen test setSensitivity (95%CI)Specificity (95%CI)AccuracyF1-ScoreAUC-ROC (95%CI)Training Time (min)†P-valueRF92% (0.8536-1)29.49% (0.1899-0.4)52.85%59.15%0.8238 (0.7394–0.8962)7.130.53XG-Boost93.33% (0.8478-1)66.23% (0.5507–0.764)76.05%74.34%0.8160 (0.7457–0.8599)12.430.4072SVM91.11% (0.8205-0.98)51.28% (0.3972–0.6203)65.58%66.13%0.6947 (0.64–0.78)2.400.000*TR-NET82.22% (0.7083–0.9184)78.21% (0.6883-8701)79.77%73.47%0.8470 (0.7669–0.9185)3.73†P-values of DeLong test: RF vs. TR; SVM vs. TR; XG-Boost vs. TRP-value < 0.05 is significant Performance metrics of predictive models for hodgkin lymphoma treatment response prediction on an unseen test set †P-values of DeLong test: RF vs. TR; SVM vs. TR; XG-Boost vs. TR P-value < 0.05 is significant Among these models, TR-NET demonstrated the most balanced sensitivity and specificity, coupled with the highest AUC-ROC, distinguishing its performance from others. SHAP analysis was applied to TR-NET to interpret these outcomes. The ROC curves for all models are provided in Supplementary Figure S1, illustrating their comparative classification performance. Explainable AI results SHAP was applied to the Tensor Radiomics Network (TR-NET), the highest-performing model (Table 4), to evaluate feature importance. Figure 5 illustrates the SHAP values generated by the TR-Net model. Positive SHAP values indicate features that increase the likelihood of predicting non-response (Class 1), whereas negative SHAP values correspond to features contributing to the prediction of treatment response (Class 0). The top features influencing TR-Net’s predictions are also highlighted, including: (1) lbp-3D-m1_ngtdm_Coarseness_CT_10, representing Neighboring Grey Tone Difference Matrix (NGTDM) coarseness from three-dimensional Local Binary Pattern (3D LBP) on CT images at a bin width of 10, with a mean SHAP value of -0.0785; (2) LHL_ngtdm_Coarseness_SUV_0.8, denoting NGTDM coarseness from low-high-low (LHL) wavelet decomposition of PET SUV at a bin width of 0.8, with a mean SHAP value of -0.0702; (3) LHH_ngtdm_Coarseness_SUV_0.6, indicating NGTDM coarseness from low-high-high (LHH) wavelet decomposition of PET SUV at a bin width of 0.6, with a mean SHAP value of -0.0334; (4) lbp-3D-m1_glcm_ClusterProminence_CT_25, reflecting Grey Level Co-occurrence Matrix (GLCM) cluster prominence from 3D LBP on CT at a bin width of 25, with a mean SHAP value of -0.0255; and (5) LLH_glcm_ClusterProminence_SUV_0.8, representing GLCM cluster prominence from low-low-high (LLH) wavelet decomposition of PET SUV at a bin width of 0.8, with a mean SHAP value of 0.0172. These features capture textural properties critical to treatment response predictions. SHAP Values for TR-NET Model Predictions. SHAP (SHapley Additive exPlanations) summary plot illustrating the most influential tensor radiomics features contributing to the TR-NET model’s prediction of treatment response in Hodgkin lymphoma (HL). In this model, positive SHAP values indicate features that increase the likelihood of predicting non-response, while negative SHAP values indicate features that contribute to predicting treatment response. The mean absolute SHAP value ranks feature importance across all samples Discussion Radiomics has shown significant potential for predicting treatment response in cancer therapy by improving treatment planning and patient outcomes. Its ability to extract detailed features reflecting tumor heterogeneity (shape, texture, intensity) captures complex biological and microenvironmental dynamics, providing valuable inputs for predictive modeling. By quantifying tumor heterogeneity, radiomics enables the identification of subtle patterns associated with treatment resistance or response, enhancing clinical decision-making. However, a key challenge is the lack of standardization, which leads to reproducibility issues and difficulties in interpreting features. Factors like patient characteristics, imaging protocols, and discretization methods influence radiomics results, hindering widespread clinical adoption [23]. Discretization, a crucial step in transforming continuous voxel intensities into discrete bins for texture matrices like GLCM, simplifies intensity values while preserving essential texture patterns. TR enhances this by considering multi-dimensional data and complex spatial relationships. The choice of bin width determines discretization and influences the extracted spatial and intensity features [24]. Furthermore, the high dimensionality of radiomics data, often coupled with limited sample sizes, necessitates robust feature selection and reduction techniques. Classic machine learning methods encounter challenges during preprocessing and feature selection. Removing correlated features, while beneficial for model performance, necessitates careful consideration to preserve biological significance. Univariate analysis, although effective for dimensionality reduction, may overlook important interactions between features. Feature reduction techniques such as principal component analysis (PCA) and recursive feature elimination (RFE) further address the challenges of high-dimensional data by reducing the feature set to those most relevant for prediction. However, PCA’s orthogonal transformations can obscure biological interpretability, while RFE may discard features involved in complex interactions. Achieving a balance between dimensionality reduction and the retention of biologically meaningful information is crucial but remains a persistent challenge. Deep learning, when integrated with radiomics, addresses many of these limitations by automating feature selection and learning hierarchical patterns. Explainable AI (XAI) models, such as SHAP, bridge the gap between model prediction and clinical utility by providing insights into individual feature contributions, enhancing clinical trust and usability. The study investigated the potential of PET/CT-based radiomics for predicting early chemotherapy response in Hodgkin lymphoma, with a specific emphasis on lesion-level analysis. Early identification of non-responding lesions is crucial for implementing risk-adapted, individualized treatment strategies. By focusing on lesion-level response, clinicians can better detect intra-patient heterogeneity in treatment sensitivity, thereby supporting more precise decisions such as intensifying therapy for resistant lesions while avoiding overtreatment in responsive areas. In this study, we implemented and evaluated a range of predictive models, including conventional machine learning algorithms (Random Forest, XGBoost, and Support Vector Machine) as well as a deep learning framework, TR-NET, to assess treatment response in Hodgkin lymphoma using tensor radiomics features extracted from PET/CT imaging. The TR-NET architecture, originally introduced by Rahmim et al. [24], was redesigned in our work to accommodate dual-modality input. While earlier applications focused primarily on single-modality data, our adaptation integrates radiomics features from both PET and CT images, with each derived at multiple discretization levels. As shown in Fig. 3, the model architecture includes seven parallel, flavor-specific subnetworks—four for PET and three for CT—each corresponding to a specific bin width. These subnetworks generate compact latent representations that capture modality- and bin-specific texture and intensity patterns. The outputs of all subnetworks are then concatenated and passed through a series of fully connected layers, forming an end-to-end learning pipeline. Tensor radiomics were extracted using various bin widths for PET and CT images. The t-SNE plot (Fig. 4) provides a two-dimensional visualization of the distribution of 1,970 radiomics features across different bin widths. However, some overlap is observed between responding and non-responding lesions; larger bin widths (SUV 0.8, CT 25, and CT 50) demonstrate improved clustering, suggesting enhanced feature differentiation. This indicates that while the predefined features capture key metabolic and structural variations, their distribution may be influenced by discretization choices and the inherent complexity of lesion characteristics. These findings emphasize the importance of optimizing feature engineering strategies to enhance class separability, increase discriminatory power, and reduce the risk of overfitting. Performance analysis (Table 4) highlights that TR-NET achieved the highest AUC-ROC (0.8470, 95% CI: 0.7669–0.9185), with a balanced sensitivity (82.22%) and specificity (78.21%), demonstrating capability in distinguishing both non-responding (class 1) and responding (class 0) lesions. While XGBoost showed a comparable AUC (0.8160), it favored sensitivity (93.33%) at the cost of lower specificity (66.23%), potentially leading to overtreatment. Similarly, RF achieved a near-identical AUC (0.8238) and high sensitivity (92%), but its very low specificity (29.49%) indicates poor performance in correctly identifying treatment responders, which could result in unnecessary therapeutic escalation. DeLong test results confirmed that TR-NET’s AUC was significantly higher than that of SVM (AUC 0.6947, p < 0.001), while the differences with XGBoost (p = 0.4072) and RF (p = 0.53) were not statistically significant, indicating similar AUC performance. TR-NET offers a more clinically appropriate balance between sensitivity and specificity, while maintaining strong sensitivity, it meaningfully enhances specificity, reduces false positives, and supports more precise treatment decisions. SHAP analysis provided additional insight into the model’s decision-making process and offered interpretability for key tensor radiomics features (Fig. 5), which may support future biomarker development. Several of the most influential features were associated with coarseness metrics derived from both PET and CT images, reflecting underlying spatial intensity patterns related to tumor heterogeneity and treatment resistance. In particular, lbp-3D-m1_ngtdm_Coarseness_CT_10 emphasized the relevance of structural homogeneity in CT at a moderate bin width (10), suggesting that more uniform anatomical textures may reflect lesions with better perfusion and less fibrosis—features associated with improved response. For PET-derived features, wavelet-transformed coarseness metrics—such as wavelet-HHH_ngtdm_Coarseness_SUV_0.8 and wavelet-HLH_ngtdm_Coarseness_SUV_0.6—were among the most predictive. These findings are consistent with prior studies [14, 43] demonstrating that metabolic heterogeneity on 18 F-FDG PET/CT is associated with poor prognosis and treatment failure in HL. In our model, positive SHAP values indicated that greater heterogeneity (i.e., lower coarseness) was linked to non-response. This aligns with the broader understanding that intratumoral heterogeneity reflects biological complexity, therapy resistance, and microenvironmental variation, all of which can compromise treatment efficacy. Notably, bin widths of 0.6 and 0.8 captured both fine- and coarse-scale patterns, respectively, striking a balance between sensitivity to localized irregularities and robustness against image noise. To our knowledge, no prior research has specifically investigated tensor radiomics in HL; therefore, we compared our findings with the radiomics approach in HL and TR studies in other cancers (Table 5). Triumbari et al. [16] analyzed lesion-specific radiomics from baseline PET/CT to improve risk stratification, focusing on the largest (Lesion_A) and hottest (Lesion_B) lesions. They found that Lesion_A’s morphological features were the strongest predictors of early treatment response (AUC: 0.78 for Deauville Score), while Lesion_B’s metabolic characteristics were more predictive of 24-month progression-free survival (PFS) (AUC: 0.74). Machado et al. [44] further explored radiomics-based prediction of treatment failure, extracting entropy, zone%, high-intensity large area emphasis, and complexity features from baseline PET/CT scans. Their machine learning model achieved an AUC of 0.96 for lesion classification, demonstrating 100% sensitivity and specificity at the patient level. Driessen et al. [21]developed a prognostic model integrating radiomic features with clinical data in relapsed/refractory HL patients. They identified metabolic tumor volume (MTV) and interlesional heterogeneity metrics (distance, volume, and uptake variability) as key predictors of three-year PFS, achieving AUCs of 0.810 and 0.750 in the training and validation cohorts, respectively. Table 5Comparison of TR-NET with prior studies utilizing tensor radiomics in cancer researchReferences2023 [16]2023 [44]2023 [21]2025 [45]2023 [25]2023 [26]2023 [47]2024 [27]PresentHL Radiomics FeatureTensor Radiomics FeaturesPatient Number22757 (Lesion: 320)182197408 (Public Dataset)408 (Public Dataset)32412870/420 LesionsAimsPredicting DS and PFS treatment response using baseline PET/CT.Predicting treatment failure using baseline 18 F-FDG PET/CT lesion and patient radiomics.Developing a 3-year PFS prognostic model for relapsed/refractory.Evaluating the predictive role of two-lesion radiomics for predicting DS and PFS, in comparison to other clinical and conventional PET/CT models.To predict PFS using RFs within a tensor radiomics frameworkComparing radiomics features (handcrafted, deep, hybrid tensor) to predict PFS.To predict OS and RFS using tensor radiomics and deep features.To develop a novel prognostic prediction model for NPC using FTR features from PET/CT imaging.To predict the treatment response for HLTensor RadiomicsN/AN/AN/AN/A17 flavours (CT-only, PET-only, and 15 fused images).17 flavours (CT-only, PET-only, and 15 fused images). 15,680 deep features29 flavours (CT-only, PET-only, and 27 fused images). and 8000 deep featureswavelet-filtered images (8 decompositions for each PET/CT image). Discretization levels (8, 16, 32, 64, 128)Discretization levels (PET: 0.2–0.8, CT: 10,25 and 50) and Multiple filtersFeature SelectionMann-WhitneyPearson correlationBackward feature selectionN/APCA and ANOVAPolynomial functionspolynomial transformsPCAANOVA. TensorLow variance PCAmRMR, ANOVAMIUnivariate analysis with Benjamini –Hochberg correction, Spearman’s correlation (threshold > 0.8)Fully automated within the TR-NET for deep learning.Low variance Correlation and ANOVA for Machin learningMethodsLogistic RegressionRandom Forest AdaBoostLogistic RegressionCox models and LASSO regressionLogistic Regression, RF, and MLPMLP, CNN, Logistic Regression, and RFGBSA, CGBSA, SVM, Random Survival Forest, and Extra Survival treesMultivariate Cox Proportional Hazard RegressionTR-NET, SVM, RF, and XGboostEvaluation Criteria & ResultsThe best AUC: 0.78The best AUC:0.96AUC: 0.750In Validation, the most effective validated model for PFS was “Clinical + Radiomics”, achieving a C-index of 66.3%. no model was successfully validated for predicting DS.Tensor radiomics + polynomial + ANOVA + LR (76.7%±3.3%) outperformed PCA fusion (70.0%±4.2%, 70.6%±6.7% external)C-indices: 0.60 ± 0.06 (OS), 0.58 ± 0.05 (RFS) using tensor Deep Features + PCA + SVM. Tensor RFs performed slightly worse.The FTR model outperformed PET and CT models in PFS prediction with a higher C-index (0.719 vs. 0.665 vs. 0.687)TR-NET achieved performance (AUC-ROC: 0.847, sensitivity: 80.2%, specificity: 78.2%) with SHAP-based interpretabilityPC: Nasopharyngeal Carcinoma ; FTR: Fusion Tensor Radiomic; RFs: Radiomics Features ; PET/CT: Positron Emission Tomography; DS: Deauville Score; PFS: Progression-Free Survival; PCA: Principle Component Analysis; RF: Random Forest; MLP: Multi-Layer Perceptron; CNN: Convolutional Neural Network; OS: Overall Survival; RFS: Recurrence-Free Survival; GBSA: Gradient Boosting Survival Analysis ; CGBSA: Component-Wise Gradient Boosting ; SVM: Support Vector MCHINE; mRMR: Minimum Redundancy Maximum Relevance ; MI: Mutual Information Comparison of TR-NET with prior studies utilizing tensor radiomics in cancer research PCA and ANOVA Polynomial functions polynomial transforms PCA ANOVA. Tensor Low variance PCA mRMR, ANOVA MI Fully automated within the TR-NET for deep learning. Low variance Correlation and ANOVA for Machin learning PC: Nasopharyngeal Carcinoma ; FTR: Fusion Tensor Radiomic; RFs: Radiomics Features ; PET/CT: Positron Emission Tomography; DS: Deauville Score; PFS: Progression-Free Survival; PCA: Principle Component Analysis; RF: Random Forest; MLP: Multi-Layer Perceptron; CNN: Convolutional Neural Network; OS: Overall Survival; RFS: Recurrence-Free Survival; GBSA: Gradient Boosting Survival Analysis ; CGBSA: Component-Wise Gradient Boosting ; SVM: Support Vector MCHINE; mRMR: Minimum Redundancy Maximum Relevance ; MI: Mutual Information To develop radiomics-based models for predicting early treatment response, it is essential to validate them on independent datasets to ensure reliability, reproducibility, and clinical relevance. Another study by Triumbari et al. [45] on classical HL patients illustrates the importance of validation in predictive modeling. Their large, retrospectively validated study aimed to evaluate the predictive role of two-lesion (Hottest and largest lesion) radiomics compared to conventional clinical and PET/CT-based models. The combined clinical and radiomics model (C + R) demonstrated the best performance for predicting progression-free survival (PFS) in validation (C-index: 66.3%). However, no model, including radiomics alone, successfully predicted early treatment response based on the Deauville Score. Ortega et al. [46] showed that integrating imaging and clinical data improves predictive modeling in non-Hodgkin lymphoma. Their model combined PET and CT radiomics with clinical factors to predict treatment outcomes. Key predictors included abnormal lymphocytes, PET entropy, and CT sphericity for overall survival, and PET/CT texture for disease-free survival. The predictive performance progressively improved from a clinical-only model (AUC 0.626) to a clinical + PET + CT model (AUC 0.759), demonstrating the additive value of imaging features. These studies support the growing role of radiomics models and underscore the necessity of validation when developing models for risk stratification and treatment outcome prediction in lymphoma. Tensor radiomics provides a robust solution by capturing tumor heterogeneity across various PET/CT bin widths, resulting in a richer and more adaptable feature space for prediction. Salmanpour et al. [25] tensor radiomics to head and neck cancer (HNC) by extracting RFs from 17 ‘flavours’; PET-only, CT-only, and 15 fused PET/CT images. Dimensionality reduction was performed using ANOVA and polynomial transforms, and logistic regression (LR) was used as the classifier, achieving an accuracy of 76.7% in cross-validation and 70.6% in external testing. Peng et al. [27] used applied Fusion Tensor Radiomics (FTR) to nasopharyngeal cancer (NPC) by extracting 4,383 features from PET and CT images (original + 8 wavelet-transformed), and generating tensor features across five discretization levels (bin widths: 8, 16, 32, 64, 128) and five wavelet fusion weights. After feature selection using univariate Cox regression and correlation filtering, an FTR model was built using multivariate Cox regression. The FTR model achieved the best performance with a C-index of 0.719. Another HNC study by Salmanpour et al. [26] expanded the TR approach by incorporating 17 flavours from PET, CT, and fused PET/CT images and tensor deep features (tensor DFs) extracted across flavours. This study aimed to assess the added value of tensor DFs in survival prediction. Their results showed that tensor DFs combined with machine learning (PCA + ANOVA + MLP) achieved superior performance, 85.3% accuracy in external testing, compared to RFs and end-to-end CNNs. A lung cancer study [47] demonstrated the effectiveness of tensor DFs and hybrid machine learning systems (HMLS) for predicting overall survival (OS) and recurrence-free survival (RFS). This study extracted RFs and deep features from 29 ‘flavours’ (PET-only, CT-only, and 27 fused PET/CT images). Advanced fusion methods and dimensionality reduction techniques (e.g., PCA and ANOVA) were employed, followed by survival prediction using Gradient Boosting, Random Forest, and Survival SVM. The best-performing models achieved c-indices up to 0.61 for OS and 0.58 for RFS, particularly when using tensor DFs combined with HMLS. While these studies demonstrate the potential of tensor radiomics for improving prognostic accuracy, their reliance on complex and time-consuming manual preprocessing limits their efficiency and scalability. TR-NET promotes collaborative, data-driven decision-making in clinical settings by providing accurate and transparent response probability predictions through a unified deep learning architecture with automated feature selection and SHAP-based interpretability. Despite the promising performance of TR-NET, this study has several important limitations. First, the absence of external validation using an independent, multi-center dataset restricts the generalizability of the model’s performance across diverse clinical settings, imaging protocols, and patient populations. Future work will prioritize external validation to confirm TR-NET’s robustness in real-world environments. Second, while SHAP analysis improves interpretability by identifying influential tensor radiomics features, it does not establish causal relationships or provide biological validation. Features such as NGTDM coarseness and GLCM cluster prominence, though statistically important, require further histopathological or molecular investigation to confirm their biological relevance and role in treatment resistance or response. Third, the high dimensionality of the feature space increases the risk of overfitting, even though we applied techniques such as SMOTE, cross-validation, and regularization. This risk may be further reduced in future studies by expanding the dataset and incorporating advanced dimensionality reduction techniques. Finally, TR-NET’s lesion-level prediction framework may limit its direct applicability to clinical decision-making, which typically relies on patient-level assessments such as overall treatment response or progression-free survival. Future research will explore strategies to aggregate lesion-level outputs into patient-level predictions and validate these against clinical endpoints to enhance translational value. Conclusion This study presents TR-NET, a deep learning model designed to predict treatment response in HL using multi-dimensional tensor radiomics features from PET and CT images. SHAP analysis was employed to explain the model’s predictions, highlighting key wavelet and texture features that influenced classification outcomes. These findings suggest that certain imaging features warrant further investigation as biomarkers. While TR-NET showed encouraging results in our internal dataset, external validation across larger, multi-center cohorts is essential before clinical application. The model’s interpretability could support future research in developing transparent and reliable decision-support tools in oncology. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Not applicable. Abbreviations Hodgkin lymphoma 2-Deoxy-2-[fluorine-18]-fluoro-D-glucose Positron emission tomography–computed tomography Standardized uptake values Radiomics features Tensor radiomics Explainable AI Shapley additive explanations Standards for reporting of diagnostic accuracy studies Checklist for artificial intelligence in medical imaging Doxorubicin hydrochloride (Adriamycin), bleomycin sulfate, vinblastine sulfate, and dacarbazine Ordered subsets expectation maximization Digital imaging and communications in medicine Generative adversarial network Image biomarker standardization initiative Grey level co-occurrence matrix Grey level run length matrix Grey level size zone matrix Grey level dependence matrix Neighboring grey tone difference matrix Laplacian of Gaussian filter t-distributed stochastic neighbor embedding Synthetic minority over-sampling technique Random forest eXtreme gradient boosting Support vector machine Deep neural network Mean squared error Receiver operating characteristic curves The area under ROC- curve Deep learning important features Principal component analysis Recursive feature elimination Nasopharyngeal carcinoma Fusion tensor radiomics Progression-free survival Head and neck cancer Deep tensor radiomics Convolutional neural network Scaled exponential linear unit Low-low-low Low-low-high Low-high-low Low-high-high High-high-high High-low-high High-high-low High-low-low Local binary pattern Inverse difference Inverse difference normalized Inverse difference moment Inverse difference moment normalized Informational measure of correlation Matthews correlation coefficient Author contributions MJ contributed to conceptualization, methodology, data curation, formal analysis, original draft writing, and visualization. HJR and ME contributed to methodology and review and editing. VR contributed to formal analysis and review and editing. FE contributed to validation and review and editing. PG contributed to conceptualization, methodology, review and editing, and supervision. SE contributed to conceptualization, review and editing, and supervision. All authors read and approved the final manuscript. Funding No Applicable. Data availability The Data supporting this article is available from the corresponding authors upon reasonable request. Declarations Ethics approval and consent to participate The Mashhad University of Medical Sciences, Mashhad, Iran, Research Ethics Committee granted ethical approval (Ethics Code: IR.MUMS.REC.1401.347). All participants provided informed consent and remained anonymous, ensuring compliance with the ethical principles outlined in the Declaration of Helsinki. Consent for publication Not Applicable. Competing interests The authors declare no competing interests or conflicts of interest. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Parham Geramifar, Email: pgeramifar@sina.tums.ac.ir. Saeid Eslami, Email: EslamiS@mums.ac.ir. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The Data supporting this article is available from the corresponding authors upon reasonable request.