Magnetic Resonance Imaging (MRI) is an essential component of current cancer treatment systems for imaging interior anatomical structures. However, accurate real-time tumour segmentation for adaptive radiation therapy is complicated by tissue motion and deformation. Precise registration is often challenging for standard imaging registration algorithms due to large displacement errors and the need for real-time application in clinical settings. This challenge is further intensified for accurate real-time tracking algorithms used in cancer therapy. For the proposed application involving real-time tumor tracking using the 2D cine-MRI technique and a deep learning approach for deformable image registration tasks with the enhanced TransMorph model, the model has been designed to include attention gates in the skip connections of the decoder. Additionally, the approach employs a two-step robust learning process: the first step involves unsupervised learning from patient image data to capture generic deformation information. To ensure boundary accuracy, a combined loss function method has been adopted, comprising boundary-weighted Dice loss, Adaptive Mean-Squared Error (MSE) loss with L1 regularizer, Smooth-Diffusion loss, and Edge loss. Validated on the multi-institutional TrackRAD2025 dataset, the proposed framework consistently outperformed baseline transformer-based models in both overlap metrics and boundary error reduction. These results highlight the potential of the attention-guided framework to enable accurate, real-time tumor tracking in MRI-guided radiotherapy.

In real-world occupational settings, mental fatigue commonly emerges from the combination of sleep deprivation with prolonged cognitive and physical workload. However, this multidimensional fatigue profile is rarely captured in controlled experimental paradigms that examine brain activation and fatigue-related responses. Consequently, the validity and transferability of cognitive fatigue biomarkers identified in passive, laboratory-based fatigue paradigms to real-world active occupational conditions remain largely unexplored. This work investigates the functional connectivity of resting-state EEG in an active (real-world) and a passive (lab-induced) mental fatigue dataset, employing a multilayer graph representation in sensor-space that integrates both within-band and cross-frequency connectivity information. Specifically, functional connections within five EEG frequency bands are quantified using Phase Lag Index (PLI), while cross-frequency interactions between bands are characterized using Phase-Amplitude Coupling (PAC). Subsequently, a dataset-specific feature extraction pipeline is adopted, utilizing multiple machine learning classifiers to co-validate the most discriminative features, while the resulting network topologies are interpreted in light of established mental fatigue neurophysiological mechanisms. Our results promote the inter-frequency PAC metric as more informative for fatigue discrimination in both datasets, compared to the intra-frequency PLI, with smaller feature sets needed in the "active" vs. "passive" fatigue setting for information saturation. The topology of the selected feature sets indicates a reorganization of neural resources that is strictly dependent on the nature of fatigue induction.

The reliability of unrelated-donor searches depends on high-resolution HLA typing, yet a large fraction of records in national stem-cell donor registries were generated at low or intermediate resolution and are therefore under-used in modern matching. Here we develop an Extreme Learning Machine (ELM) approach that upgrades low/mid- to high-resolution HLA data by learning the haplotype and diplotype structure of a national donor population and assigning the most probable high-resolution genotypes together with posterior probabilities. The model was trained on the Greek national registry (Hellenic Transplant Organization, established 2002; 117,345 donors, ~20% low-resolution) and validated on two independent Greek cohorts (ORAM, n = 20,100; GRPT, n = 4,353) using accuracy and call-rate metrics. The population-specific ELM achieved a per-locus accuracy of 70–94% (depending on the confidence threshold) with an overall call rate of 98.1%, recovering usable high-resolution information and increasing the proportion of registry donors usable in high-resolution matching. The method is fast, lightweight and population-tailored, complementing established expectation-maximisation imputation tools.

Developmental dysplasia of the hip (DDH) is a common musculoskeletal disorder in infancy, and early detection is essential for optimal clinical outcomes. Radiographic assessment is traditionally based on angular measurements, which may be limited by variability in landmark identification and do not fully capture the complex morphology of the hip joint. In this study, we investigate whether geometric features derived from the hip joint articulation space can be used to differentiate between normal and dysplastic hips in infant radiographs. Pelvic X-ray images from infants (mean age 4.5 ± 0.83 months) were analyzed, and custom segmentation masks were developed to isolate the joint space region. A total of 99 geometric and radiomic features were extracted and evaluated using statistical analysis and supervised machine learning methods. Multiple features demonstrated strong discriminative power between normal and DDH (p < 0.001), with shape and spatial distribution characteristics showing the highest relevance. Classification models achieved an F1-score of approximately 80% on the full dataset. Notably, patient age was identified as a significant confounding factor, and analysis on an age-matched subset improved classification performance to 94% accuracy and 93% recall. These findings suggest that geometric characterization of the hip joint space provides a promising and interpretable framework for DDH detection. The results also highlight the importance of age-stratified analysis in pediatric imaging. Further validation on larger and more diverse datasets is required to assess clinical applicability.

Radiotherapy workflows for oncological patients increasingly rely on multi-modal medical imaging, commonly involving both Magnetic Resonance Imaging (MRI) and Computed Tomography (CT). MRI-only treatment planning has emerged as an attractive alternative, as it reduces patient exposure to ionizing radiation and avoids errors introduced by inter-modality registration. While nnU-Net-based frameworks are predominantly used for MRI-to-CT synthesis, we explore Mamba-based architectures for this task, aiming to showcase the advantages of state-space modeling for cross-modality translation compared to standard convolutional neural networks. Specifically, we adapt both the U-Mamba and the SegMamba architecture, originally proposed for segmentation, to perform cross-modality image generation. Our 3D Mamba architecture effectively captures complex volumetric features and long-range dependencies, thus allowing accurate CT synthesis while maintaining fast inference times. Experiments were conducted on a subset of SynthRAD2025 dataset, comprising registered single-channel MRI-CT volume pairs across three anatomical regions. Quantitative evaluation is performed via a combination of image similarity metrics computed in Hounsefield Units (HU) and segmentation-based metrics obtained from TotalSegmentator to ensure geometric consistency is preserved. The findings pave the way for the integration of state-space models into radiotherapy workflows.

Intrusive saccades during active visual fixation indicate deficits in inhibitory control which is crucial for cognitive control function. Research has shown that abnormalities in these mechanisms are linked to neurological disorders such as schizophrenia and obsessive-compulsive disorder (OCD), both involving dysfunctions in frontal-subcortical circuits. Eye movement studies and machine learning (ML) techniques have been used to differentiate clinical from neurotypical populations. This study aimed to classify healthy controls, patients with OCD and schizophrenia patients, based on oculomotor behavior during active fixation tasks and provide insights into related neurophysiological mechanisms. Data from three visual fixation tasks were analyzed using statistical tests to select saccade features to be used in the classification. A shallow Artificial Neural Network (ANN) was implemented for binary and three-class classification. Binary classification achieved 87% accuracy and 93% specificity in distinguishing controls from the patients with schizophrenia group, 84% accuracy and 90% sensitivity in distinguishing between controls and medicated patients with OCD not taking antipsychotics, while differentiation between patients with schizophrenia and medicated patients with OCD not taking antipsychotics reached 77% accuracy and 82% specificity. The findings provided indications that selected saccadic features can differentiate OCD and schizophrenia patients from healthy controls using shallow ANNs, while distinguishing between OCD and schizophrenia patients remains more challenging. Notably, tentative indications were provided that group differences were driven more by intrinsic saccadic generation properties than by fixation or inhibitory mechanisms, concerning unwanted saccades that are intrusive in nature in the context of fixation.

Clinical question answering over electronic health records (EHRs) increasingly relies on large language model (LLM) agents that retrieve structured patient data through external tools. Published benchmarks, however, evaluate these systems at a single patient-population size, and rarely measure the effect of backend representation from that of the retrieval interface design. This paper compares six retrieval configurations that vary along two axes: backend (a property graph database, a relational database and a dense vector index) and interface design (curated domain-specific tool calls, model-generated queries, full-text search, and single-shot dense retrieval). The evaluation covers a 334-question bank spanning six categories (simple lookup, multi-hop, temporal, cohort, reasoning, and unanswerable), instantiated at three nested population scales: 200, 2000, and 20,000 alive patients from a single Synthea cohort. Four models are compared: Claude Haiku 4.5, Qwen 2.5 72B, Llama 3.1 8B, and Llama 3.3 70B, spanning closed-frontier and open-source alternatives. Curated tool-calling configurations improve accuracy over retrieval-augmented baselines for capable models, but reduce accuracy for a small open-source model due to function-calling protocol failures. We report how accuracy, latency, and cost evolve with each approach, model size, and cohort size, supported by paired statistical tests and confidence intervals. All benchmark components, databases, and evaluation code are publicly available.

Heart failure management increasingly relies on heterogeneous clinical and real-world data generated through remote monitoring technologies. However, transforming these multimodal data streams into actionable insights requires robust interoperability infrastructures. This study presents the RETENTION interoperability framework, a federated HL7 Fast Healthcare Interoperability Resources (FHIR)-based architecture designed to support multi-site heart failure monitoring across five independent clinical environments. A semantic reference model comprising 444 clinical and contextual variables was developed and aligned with FHIR R4 resources and internationally recognised terminology systems. The platform adopts a selective profiling strategy, extending only the Patient resource while standardising the remaining variables through example-driven Implementation Guide documentation. Identifiable data are retained locally within Clinical Site Backends, whereas anonymised datasets are periodically aggregated into a Global Insights Cloud to enable centralised analytics and controlled third-party interactions. The framework was deployed across six hospitals (with two Spanish hospitals sharing the same deployment), supporting 390 patients and over 130,000 patient-days of monitoring, with more than 3.6 million remote device data points harmonised without schema conflicts. The results demonstrate that large-scale semantic harmonisation and privacy-preserving aggregation can be achieved using a lightweight profiling approach, providing a scalable and reproducible interoperability model for multi-centre digital health research infrastructures.

Magnetic resonance imaging (MRI)-guided radiotherapy (MRIgRT) integrates MRI with linear accelerators (MRI-linacs), enabling real-time motion management based on temporally resolved 2D MRI (cine-MRI). Current systems rely on template matching or deformable image registration for radiotherapy target (typically the gross tumor volume) localization, which allows beam gating. Further advances in localization could support more precise and efficient delivery methods. https://trackrad2025.grand-challenge.org/ was organized to provide a common dataset to benchmark algorithms for MRIgRT target tracking in 2D+t cine-MRI. Participants propagated target segmentation masks from an initialization frame across subsequent frames. The dataset comprised sagittal cine-MRI scans of 585 cancer patients undergoing radiotherapy at 0.35 T and 1.5 T MRI-linacs at six different institutions, with expert-annotated targets in 108 sequences. Target sites included the thorax (179 cases), abdomen (266 cases), and pelvis (140 cases). A total of 477 unlabeled and 50 labeled cases were provided for training purposes, 58 cases were kept private for preliminary testing (8) and final evaluation (50). The algorithms submitted by participants were executed on the challenge platform and assessed using metrics in three categories: geometric accuracy, surrogate dose accuracy and execution speed. Rankings were derived via a Rank-Then-Mean scheme. TrackRAD2025 attracted 148 registrations from 28 countries, 100 preliminary submissions and 24 final submissions from 14 teams. The top five methods achieved mean Dice similarity coefficients >0.87 and Euclidean center distances <2.1 mm, comparable to interobserver variability. Leading top five solutions featured foundation models with (4) or without (1) finetuning. Field strength had minimal effect on performance and tracking worked better for the pelvis with reduced motion amplitude compared to the thorax and abdomen cases, which achieved equivalent performance. TrackRAD2025 established a benchmark for MRIgRT tracking on multi-institutional cine-MRI data, highlighting foundation models as promising for clinical translation.

Accurate segmentation of abdominal organs in Computed Tomography (CT) underpins radiotherapy planning, surgical planning, and disease monitoring. Existing benchmarks rank architectures by a single aggregate Dice score, without per-organ statistical testing or boundary-sensitive metrics, even though models are chosen organ by organ for clinical use. We benchmark ten architectures spanning convolutional, attention-based, transformer, and state–space (Mamba) families on the AMOS CT dataset under one identical nnU-Net-style pipeline; we report per-organ Dice, 95-percentile Hausdorff Distance (HD95), and Normalised Surface Dice, with pairwise significance tested on an independent external dataset (TotalSegmentator). A competitive cluster of convolutional and Mamba models leads; rankings are stable on large organs but reshuffle by 10–13% on the small, geometrically complex ones, and boundary fidelity separates the models into tiers that the Dice ranking hides. This ordering largely holds on the external set (Spearman ρ=0.84). Selecting a model on aggregate Dice alone is therefore unsafe for organ-specific clinical tasks: per-organ overlap and boundary metrics should be the primary acceptance criteria for selecting a model before clinical deployment.

BACKGROUND:Epicutaneous patch testing is the gold standard for diagnosing allergic contact dermatitis (ACD), yet its interpretation relies on subjective scoring and remains prone to inter-observer variability. AIM:In this study, we present a machine learning pipeline that complements subjective scoring with objective bioengineering measurements derived from the Antera 3D imaging system. METHODS:A dataset of skin reactions was analyzed, with particular attention to how the data was split to avoid information leakage between patients. For this reason, methods such as GroupShuffleSplit and GroupKFold, which account for patient-level clustering, were used. RESULTS:Of the models tested, the Random Forest classifier showed the best overall performance, with an AUC of 0.861 (95% CI: 0.830-0.888) on patient data that had not been used during training, outperforming the Multi-Layer Perceptron model. The incorporation of additional features that capture changes between 48 and 72 h improved the results even further, raising the AUC to 0.902 and achieving a very high sensitivity of 96.8%. CONCLUSIONS:Overall, the results show that objective biophysical measurements derived from the Antera 3D imaging system can be combined with machine-learning techniques for objective patch-test assessment. Incorporating temporal changes between the 48- and 72-h readings further improved model performance, suggesting that temporal changes provide additional information beyond single-time-point measurements.

Accurate delineation of pancreatic tumors on Magnetic Resonance Imaging (MRI) is important for diagnosis, radiotherapy treatment planning, and outcome assessment, but remains challenging due to complex anatomy and subtle tumor appearance. In routine practice, tumor contours on MRI are produced manually, which is time-consuming and subject to inter-observer variability. Radiotherapy on MRI-Linear Accelerator (MRI-Linac) systems further requires fast and consistent Gross Tumor Volume (GTV) contours for online adaptation, yet most public pancreas tumor segmentation benchmarks focus on Computed Tomography (CT). The Pancreatic Tumor Segmentation in Therapeutic and Diagnostic MRI (PANTHER) challenge addresses this gap by benchmarking automatic pancreatic tumor segmentation on MRI. The dataset includes contrast-enhanced T1-weighted diagnostic MRI and T2-weighted MRI-Linac scans with expert pancreas and tumor annotations, organized into two tasks: (1) tumor segmentation on diagnostic MRI and (2) tumor segmentation on MRI-Linac images. Performance was evaluated using overlap metrics, distance-based metrics, and tumor volume error. The challenge attracted 285 registered participants, with 12 and 9 final submissions for Tasks 1 and 2, respectively. On diagnostic MRI, top methods achieved performance close to inter-reader agreement. Multi-reader analysis suggested that models often reproduced the contouring style of the training annotator, highlighting the importance of annotation quality and consensus. In contrast, performance on MRI-Linac images was lower and more heterogeneous, including cases of complete localization failure. PANTHER provides the first public benchmark for pancreatic tumor segmentation on MRI, showing that clinically useful automation is feasible on diagnostic MRI, while robust MRI-Linac GTV segmentation remains an open challenge.

This report describes our submission to the Big Cross-Modal Attenuation Correction (BIC-MAC) 2026 Challenge for CT-less PET attenuation correction through multimodal pseudo-CT synthesis. We build upon a standard nnU-Net architecture and combine anatomical and physical supervision to improve both pseudo-CT quality and downstream PET reconstruction. Anatomical supervision is introduced through a frozen TotalSegmentator feature extractor, anatomy-guided structural constraints and patch sampling, while physical supervision is achieved using a differentiable attenuation correction factor projection loss based on multi-angle attenuation projections. Furthermore, the network is initialized with pretrained weights obtained from training on the SynthRAD Challenge MR-to-CT dataset. Minimal architectural modifications are applied, while performance improvements are pursued across the nnU-Net pipeline, including preprocessing, plans, and supervision design, among other components. Our final submission demonstrates the effectiveness of combining anatomical supervision, attenuation physics, and efficient nnU-Net scaling for CT-less PET attenuation correction.

To investigate the prognostic value of an artificial intelligence (AI)–based semi-automated tool for longitudinal whole-body quantification of total metabolic tumor volume (TMTV) and total lesion glycolysis (TLG) on [1⁸F]FDG PET/CT in patients with metastatic melanoma undergoing immune checkpoint inhibitor (ICI) therapy, and to assess its prognostic relevance alongside established PET-based metabolic response criteria. Forty-three patients with unresectable metastatic melanoma treated with ICIs underwent [1⁸F]FDG PET/CT at baseline (n = 43), after two cycles (interim; n = 41), and after four cycles of therapy (late; n = 43). Whole-body tumor segmentation was performed using a previously validated AI-based framework combining ensemble unsupervised segmentation and deep representation learning, followed by expert review. TMTV and TLG were calculated for each time point. Overall survival (OS) was analyzed using Kaplan–Meier estimates, log-rank tests, and Cox proportional hazards regression. Multivariable models included LDH, AJCC stage, and ECOG performance status. Metabolic response was additionally assessed using EORTC, PERCIST, PERCIMT, imPERCIST5, and iPERCIST criteria. Median follow up [95

Accurate and real-time observation of the surgeon’s hands in real time is a fundamental component in the development of computer-assisted surgical procedures. However, challenges such as instrument occlusion, motion blur and the use of surgical gloves require effective and specialized solutions. This study presents the methodology behind the second-placed solution for Task 2 of the Trauma THOMPSON Challenge 2025. The approach is defined by a comprehensive data-driven workflow that begins with extensive Exploratory Data Analysis (EDA) to understand the unique characteristics of the available data. The results of the analysis directly informed the development of a customized data pre-processing pipeline, as well as the tuning of hyperparameters for an object detection method based on YOLOv8. By systematically addressing issues of redundancy between video frame sequences and scale variations in the objects to be recognized, this solution achieves a good balance between speed and accuracy. The validity of the pipeline has been demonstrated on the Trauma THOMPSON dataset, achieving a Mean Average Precision at Intersection over Union threshold of 0.5 (mAP50) of 73.25 https://github.com/KostasGeorgas97/The-Trauma-THOMPSON-Challenge-2025-Task-2-Hand-Tracking- .

Sleep staging is essential for evaluating sleep quality, yet manual methods remain labor-intensive and subjective. Existing automated approaches largely overlook the influence of gender and age, limiting their accuracy. We propose a cuckoo search-optimized deep forest model (CS-DF) that integrates multi-granularity scanning for diverse EEG feature extraction and incorporates gender- and age-based grouping strategies to capture demographic-specific physiological characteristics. Using the Sleep-EDFx dataset with subject-wise 10-fold cross-validation, four types of EEG features were extracted from 78 subjects grouped by gender (male/female) and age (young/middle-aged/old). The CS-DF model achieved an accuracy of 84.2%, macro-F1 of 79.3%, and kappa of 77.2%, outperforming the baseline deep forest. Under leave-one-subject-out cross-validation, gender grouping improved accuracy and kappa by up to 1.10% and 3.70%, respectively, while age grouping yielded up to 4.30% accuracy gain in the young group. The combined grouping strategy achieved overall improvements of 1.10%, 1.00%, and 3.50% in accuracy, macro-F1, and kappa. External validation on the independent ISRUC dataset confirmed the generalizability of the framework, with the combined strategy achieving an accuracy of 84.1%, macro-F1 of 81.1%, and kappa of 79.9%. These results demonstrate that integrating demographic grouping strategies within the CS-DF framework substantially enhances automated sleep staging across diverse cohorts.

Magnetic Resonance Imaging (MRI) is a key modality in cancer treatment, providing high soft tissue contrast for the visualization of tumors and internal anatomy. Radiotherapy, which is widely used in treatments, requires precise tumor segmentation to ensure targeting the true tumor and minimizing radiation exposure to healthy tissues. In this regard, real-time automatic tumor tracking from cine-MRI can provide accurate tumor localization supporting adaptive radiotherapy. Conventional image registration techniques exhibit limitations when handling large misalignments and high computational demands, unlike deep learning methods, with high learning capabilities and fast inference times. A real-time tumor tracking approach for 2D cine-MRI using deep learning-based deformable image registration, based on an improved TransMorph architecture, is presented. The approach adheres to a two-step training paradigm: (1) unsupervised pretraining on unlabeled patient image pairs, and (2) supervised fine-tuning with segmentation labels. Attention gates are integrated into skip connections to enhance spatial selectivity regarding the most relevant regions for alignment. A composite loss function is utilized, synthesizing boundary-weighted Dice, adaptive MSE, L1 and smooth diffusion and edge-based regularization. Both overlap and distance-based metrics were computed to assess the model’s segmentation accuracy in the registration of various frames within the patients. The proposed model achieved DSC 93.42%± 5% , 95HD 2.75 ± 3.24 mm, 50HD 0.88 ± 0.40 mm. Extensive benchmarking demonstrated that the proposed framework consistently achieved superior performance compared to the TransMorph model, its variants, and other existing state-of-the-art image registration approaches. Experimental results indicate that this technique may serve as a critical tool for the advancement of MRI-guided radiotherapy.

BIO-STREAMS demonstrates how large-scale, organized biobanks can revolutionize childhood obesity research while maintaining strict GDPR compliance. This paper presents a scalable infrastructure that successfully manages large amounts of sensitive health data, from over $\mathbf{5 3, 8 0 0}$ participants across six European countries (42.400 retrospective participants from existing cohorts and 11,400 prospective participants from new clinical studies). The platform demonstrates its capabilities in handling high-volume, high-velocity data streams involving clinical biospecimens, digital biomarkers, behavioral patterns, and real-time mobile application data. The system processes large amounts of multi-modal health data, leveraging federated big data architecture, privacy-preserving analytics and AI-driven machine learning pipelines. The infrastructure supports real-time stream processing, automated data harmonization at scale, and synthetic data generation using machine learning models. Initial validation demonstrates the platform's capacity to handle unprecedented data volumes while enabling cross-border big data analytics for obesity research. BIO-STREAMS establishes a new paradigm for big data applications in pediatric healthcare, showcasing scalable solutions for managing sensitive health data at European scale. It also serves as a model for digital biobanks targeting complex chronic diseases requiring multidisciplinary research.

Background: Ulcerative colitis (UC) is a chronic inflammatory bowel disease characterized by continuous inflammation of the colon and rectum. Accurate disease assessment is essential for effective treatment, with endoscopic evaluation, particularly the Mayo Endoscopic Score (MES), serving as a key diagnostic tool. However, MES measurement can be subjective and inconsistent, leading to variability in treatment decisions. Deep learning approaches have shown promise in providing more objective and standardized assessments of UC severity. Methods: This study utilized publicly available endoscopic images of UC patients to analyze and compare the performance of state-of-the-art deep neural networks for automated MES classification. Several state-of-the-art architectures were tested to determine the most effective model for grading disease severity. The F1 score, accuracy, recall, and precision were calculated for all models, and statistical analysis was conducted to verify statistically significant differences between the networks. Results: VGG19 was found to be the best-performing network, achieving a QWK score of 0.876 and a macro-averaged F1 score of 0.7528 across all classes. However, the performance differences among the top-performing models were very small suggesting that selection should depend on specific deployment requirements. Conclusions: This study demonstrates that multiple state-of-the-art deep neural network architectures could automate UC severity classification. Simpler architectures were found to achieve competitive results with larger models, challenging the assumption that larger networks necessarily provide better clinical outcomes.

Small-cell lung cancer (SCLC) is an aggressive malignancy that exhibits high prevalence for brain metastases. Furthermore, chemotherapy and metastasis-preventive approaches are also linked to neurotoxicity, further aggravating cognitive impairment. Despite evidence supporting structural and functional brain alterations in SCLC, the application of machine learning (ML) to new connectivity biomarkers has remained unexplored. This study is—to the best of our knowledge—the first to apply ML to structural brain connectomics in SCLC, using diffusion tensor imaging (DTI) to identify features discriminating between post-chemotherapy SCLC patients and healthy controls. Specifically, we constructed structural networks via deterministic tractography, applying an adapted feature reduction technique to identify the most informative connections without selection bias. This process isolated 16 connections involving 26 brain regions, predominantly in the frontal, temporal, and parietal lobes, showcasing primarily intra-hemispheric and left-lateralized alterations. Our optimal model leveraged a Gaussian Support Vector Machine (SVM), achieving a weighted accuracy of 0.92, a sensitivity of 0.93, a specificity of 0.91, and an area under the curve of 0.94. The selected feature subset retained high performance when tested with other classifiers, confirming its robustness. Our findings differ from prior studies based on statistically derived features, highlighting the ML-driven connectomics’ potential in uncovering DTI-derived SCLC patterns, offering interpretable insights for neuroimaging-based diagnostics.