From Classification to Clinical Trust: 3D Deep Learning and Layer-Wise Explainability for Brain Tumor MRI

Preoperative differentiation of glioma from meningioma on MRI is important for treatment planning but depends on expert interpretation. A 3D ResNet-34 with CBAM attention was trained on four-modality MRI from 2251 BraTS 2023 patients (1801 for four-fold cross-validation; 450 held out for testing) and achieved 96% accuracy, an AUC of 0.993, and an expected calibration error of 0.0165 on the test set. For layer-wise explanation, Grad-CAM and SmoothGrad were applied at the four residual stages, and a reconstruction decoder was added to obtain full-resolution, decoder-mediated attribution at the 4 × 4 × 4 bottleneck, where standard Grad-CAM is too coarse. Four quantitative metrics gave complementary results: the decoder-assisted attribution had the highest pointing-game accuracy (80.2%) and the lowest deletion AUC, whereas Layer 1 had the highest insertion AUC and Layer 2 the highest mean faithfulness delta, with no significant faithfulness difference among Layers 1, 2, and 4. Glioma was localized by most layers, while meningioma was localized mainly through the decoder-assisted pathway. A preliminary clinical review and a normalization analysis are also reported. The decoder offers an annotation-free way to obtain full-resolution attribution from deep layers of volumetric models.