Positive-gradient-weighted object activation mapping: visual explanation of object detector towards precise...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate polyp detection and localization are critical for colonoscopy diagnosis.
• Computer-aided diagnosis systems can prevent missed polyps and mislocalizations.

Purpose of the Study:

• To develop novel visual explaining methods for trained object detectors.
• To achieve fast, accurate polyp detection and precise automated polyp localization using bounding boxes.

Main Methods:

• Refined gradient-weighted class activation mapping (Grad-CAM) for highlighting important patterns in convolutional neural networks.
• Object activation mapping (OAM) for visualizing salient object patterns in images for detection.
• Polyp activation mapping (PAM) integrating adaptive local thresholding for precise polyp localization.

Main Results:

• The refined mapping improved visualization of important patterns compared to original Grad-CAM.
• Object activation mapping effectively visualized key patterns in colonoscopic images for polyp detection.
• Polyp activation mapping achieved mean Dice scores of 0.76, 0.72, and 0.72 on the ETIS-Larib, CVC-Clinic, and Kvasir-SEG datasets, respectively.

Conclusions:

• New visual explaining methods were developed by refining and extending Grad-CAM for convolutional neural networks.
• Experimental results validated the methods, demonstrating accurate visualization and polyp localization in colonoscopic images.
• The proposed methods aid in interpreting and applying trained polyp detectors for clinical use.