ProtoRadNet: Prototypical patches of Convolutional Features for Radiology Image Classification Network.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) excel at radiology image classification but often function as "black boxes," hindering trust and adoption by medical professionals.
• Existing patch-based prototypical networks for interpretability have seen limited application in the radiology domain.

Purpose of the Study:

• To introduce ProtoRadNet, a novel network for radiology image classification that enhances interpretability by visualizing identified prototypes.
• To refine CNN training by focusing on significant prototypes within and across classes, rather than all convolutional features indiscriminately.

Main Methods:

• ProtoRadNet integrates inter-class and intra-class prototypes to achieve both localized and global interpretability.
• The model leverages image-level ground truths, making it suitable for real-world applications where detailed annotations are scarce.

Main Results:

• ProtoRadNet demonstrates superior performance compared to state-of-the-art methods on Brain MRI, Chest CT, and MIMIC CXR-LT datasets.
• Achieved macro-averaged F1-scores of 92.16%, 96.14%, and 29.32%, with notable improvements over competing methods.

Conclusions:

• ProtoRadNet effectively bridges the gap between CNN findings and expert understanding in medical imaging.
• The model offers transparent reasoning for classification decisions, enhancing the practical utility of AI in radiology.