Fig. 2: Visual ablation analysis with Grad-CAM. | npj Digital Medicine

Fig. 2: Visual ablation analysis with Grad-CAM.

From: PrysmNet a polyp refining system using salience and multimodal guidance for reproducible cross domain segmentation

Fig. 2: Visual ablation analysis with Grad-CAM.

a–c show Grad-CAM attention heatmaps for three configurations: a baseline (ViT backbone + U-shaped decoder; without BSM/FMDM/MGM), b baseline + biologically-inspired salience module (BSM), and c the full PrysmNet with BSM plus foundation-model distillation (FMDM) and multi-modal guidance (MGM). d–f show the corresponding segmentation masks. Adding BSM sharpens boundary focus; adding FMDM and MGM—used only during training—further consolidates localization, yielding cleaner contours and more accurate masks, especially for diminutive or weak-boundary polyps; the full model (c, f) performs best.

Back to article page