Open-Source 3D Slicer Modules for CT Image Analysis: L3 Muscle Segmentation and Explainable Pulmonary Lesion Classification

Deep-learning models for computed tomography (CT) analysis are usually released as training scripts and weight files rather than as tools a clinician can open and run, and many of them assume a GPU, a network connection or a commercial licence. We describe two open-source scripted Python modules for the 3D Slicer platform that address this gap for two distinct CT analysis tasks. The first module segments skeletal muscle, the only tissue class annotated and validated in this work, on a single axial CT section at the level of the third lumbar vertebra (L3) using a 2D U-Net trained with MONAI, and automatically derives skeletal muscle area (SMA) and skeletal muscle index (SMI) and flags sarcopenia status against the sex-specific cut-offs of Prado et al.; when no trained model is available the module falls back to a coarse Hounsfield-threshold segmentation that must be corrected manually, so it always returns an editable result. The second module classifies pulmonary CT images as healthy, benign or malignant through a two-stage hierarchical ResNet-50 pipeline with a user-adjustable benign decision threshold, and displays a Grad-CAM heat map alongside every prediction. Both modules run entirely locally, on CPU if necessary, inside 3D Slicer. On an in-house dataset of 51 cases (one section per patient; 40 training and 11 validation) the segmentation model reached a mean validation Dice of 0.837 during training; on a standalone evaluation over all 11 validation cases the skeletal muscle Dice was 0.832 ± 0.271 (mean ± SD; 95% bootstrap CI 0.664–0.930), the spread being driven by a single case (Dice 0.023) whose reference mask is misregistered; secondary sensitivity analysis excluding this case gave 0.913 ± 0.042 (n = 10). On the public IQ-OTH/NCCD dataset (1097 images), Stage 1 (healthy vs. tumour) reached a macro F1 of 0.848 and Stage 2 (benign vs. malignant) a macro F1 of 0.691 with a benign recall of only 0.29; end-to-end accuracy on the 219-image validation subset was 82.6% but balanced accuracy only 63.5%, with 95.5% of malignant and 86.7% of healthy images correctly identified but only 8.3% (2/24) of benign images. We report this benign-class failure explicitly: it persisted through weighted sampling, focal loss, class-weight boosting and class-specific augmentation, and we regard it as the most informative empirical result of the work. The classification module is therefore not suitable for characterising a lesion as benign; at most it can flag images for further review. Because the public dataset provides no patient identifiers, its split was made at image level, and 198 of the 219 validation images have a near-identical image in the training subset, so these classification figures are upper bounds rather than estimates of performance on new patients. These figures reflect technical evaluation on retrospective data only; neither module has undergone clinical validation, and both are released as research tools rather than for diagnostic use.