Consistency and stability benchmarking of Grad-CAM, SHAP, and LIME for diffuse and focal brain MRI classification
Bibliographic record
Abstract
Clinical adoption of deep convolutional neural networks for brain-MRI interpretation hinges on reliable visual explanations, yet systematic head-to-head evaluations that quantify explanation robustness and reproducibility across multiple neuropathologies within a fully standardized pipeline remain scarce. This study performs a controlled comparison of Gradient-weighted Class Activation Mapping (Grad-CAM), Deep-SHAP and LIME using a single 3-D ResNet-50 backbone fine-tuned on two public datasets: Augmented Alzheimer MRI v2 (6 400 T1-weighted slices, four dementia stages) and Brain-Tumor MRI (3 264 slices, four tumor classes). Uniform preprocessing—HD-BET skull stripping, MNI-152 registration, 1 mm<sup>3</sup> resampling and z-score normalization—minimizes scanner bias, while patient-stratified 70 : 15 : 15 splits prevent information leakage. Diagnostic performance surpasses 98 % accuracy and 0.99 AUC on both tasks, providing a robust foundation for explanation assessment. These measures aim to capture how reproducible and robust the explanations are under realistic perturbations. Grad-CAM exhibits the highest consistency (0.764 for Alzheimer’s, >0.90 for tumors) and stability (0.684 and >0.90), accurately highlighting hippocampal atrophy and tumor cores. Deep-SHAP delivers anatomically detailed but computationally intensive attributions, whereas LIME shows pronounced variability due to super-voxel segmentation. Findings recommend gradient-based explanations for rapid, focal-lesion screening and reserve SHAP for diffuse pathology requiring voxel-level justification, while cautioning that LIME demands careful parameter tuning before clinical deployment. An open-source evaluation script accompanies the study to facilitate reproducible, pathology-aware XAI benchmarking in neuro-radiology. To assist related work, the code is available at: https://github.com/fyyyaug/medical-image.git.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.000 | 0.001 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".