Add Grad-CAM saliency maps for 6 CNN-class models

.gitattributes

@@ -53,3 +53,9 @@ analysis/roc_curves/roc_resnet50.png filter=lfs diff=lfs merge=lfs -text
 analysis/roc_curves/roc_retfound.png filter=lfs diff=lfs merge=lfs -text
 analysis/roc_curves/roc_swin_b.png filter=lfs diff=lfs merge=lfs -text
 analysis/roc_curves/roc_vgg19.png filter=lfs diff=lfs merge=lfs -text
+gradcam/gradcam_clip_openai.png filter=lfs diff=lfs merge=lfs -text
+gradcam/gradcam_densenet121.png filter=lfs diff=lfs merge=lfs -text
+gradcam/gradcam_inception_v3.png filter=lfs diff=lfs merge=lfs -text
+gradcam/gradcam_resnet101.png filter=lfs diff=lfs merge=lfs -text
+gradcam/gradcam_resnet50.png filter=lfs diff=lfs merge=lfs -text
+gradcam/gradcam_vgg19.png filter=lfs diff=lfs merge=lfs -text

gradcam/README.md

@@ -0,0 +1,20 @@
+# Grad-CAM Saliency Maps
+
+Per-class Grad-CAM visualisations for the 6 CNN-class models. For each model the figure
+shows the 10 disease classes (rows) with the original fundus image (left) and the
+Grad-CAM heatmap overlay (right).
+
+## Generation
+- Library: `pytorch-grad-cam`
+- Target layer: last convolutional block of each backbone
+- CLIP uses input-gradient saliency (no convolutional layer at the same semantic level)
+- Foundation transformer models (DINOv2-L, RETFound, Swin-B) require a different attention
+  rollout / hookable layer and are not included in this v1 Grad-CAM batch.
+
+## Files
+- `gradcam_vgg19.png`
+- `gradcam_resnet50.png`
+- `gradcam_resnet101.png`
+- `gradcam_densenet121.png`
+- `gradcam_inception_v3.png`
+- `gradcam_clip_openai.png`