Optimal CAM variant for treatment outcome prediction

Determine which Class Activation Mapping variant—Grad-CAM, Grad-CAM++, or pixel-wise CAM—achieves optimal performance for predicting whether a brain metastasis will exhibit a ≥20% volume reduction at 3-month follow-up after PULSAR treatment, when used within the convolutional autoencoder–classifier framework trained on T1-weighted contrast-enhanced MRI of brain metastases.

Background

The paper evaluates three CAM variants (Grad-CAM, Grad-CAM++, and pixel-wise CAM) applied to the last convolutional layer of a CNN autoencoder–classifier to interpret predictive features for response vs. non-response in brain metastases treated with PULSAR. Pixel-wise CAM provided more fine-grained localization and achieved the highest classification metrics in this cohort, but also showed potential sensitivity to noise and input perturbations.

Despite preliminary findings favoring pixel-wise CAM, the authors note that differences among CAM methods and their suitability for tumor imaging vs. natural images necessitate further investigation, leaving the question of which CAM approach is optimal for treatment outcome prediction unresolved.