ProtoRadNet: Prototypical patches of Convolutional Features for Radiology Image Classification Network.
Authors
Affiliations (3)
Affiliations (3)
- Department of Data Science and Engineering, Indian Institute of Science Education and Research, Bhopal, Bhopal Bypass Road, Bhauri, Bhopal, 462066, Madhya Pradesh, India. Electronic address: [email protected].
- Jawaharlal Nehru Cancer Hospital and Research Centre, Idgah Hills, Bhopal, 462001, Madhya Pradesh, India. Electronic address: [email protected].
- Department of Data Science and Engineering, Indian Institute of Science Education and Research, Bhopal, Bhopal Bypass Road, Bhauri, Bhopal, 462066, Madhya Pradesh, India. Electronic address: [email protected].
Abstract
Convolutional Neural Networks (CNNs) have achieved significant success in classifying radiology images; however, their implementation often resembles a "black box," limiting medical practitioners' ability to comprehend and trust the decisions made due to a lack of interpretability. Recent advancements in patch-based prototypical networks have sought to enhance the interpretability of image classification systems. Still, the use of these models, specifically developed for the radiology domain, has been limited. This paper presents ProtoRadNet - Prototypical Patches of Convolutional Features for Radiology Image Classification Network. ProtoRadNet provides explicit visualisations of the prototypes identified within an image during classification tasks, thereby offering transparent reasoning for its decisions and effectively bridging the divide between CNN findings and their practical implications to the domain experts. The primary objective of ProtoRadNet is to identify significant prototypes of convolutional features within individual classes and across all classes, refining the CNN's training to bolster interpretability rather than relying on all convolutional features indiscriminately. The model achieves localised and global interpretability by integrating inter-class and intra-class prototypes, enhancing overall decision-making processes. This interpretability is particularly noteworthy as it is accomplished using only image-level ground truths, rendering it semantically meaningful for real-world applications, where detailed annotations are frequently unavailable or time-consuming. Empirical evaluation demonstrates that ProtoRadNet surpasses state-of-the-art in most cases. It achieves macro-averaged F1-scores of 92.16%,96.14% and 29.32% with an improvement of +2.04%,+0.73% and +0.41% respectively than the best competing method on Brain MRI, Chest CT and MIMIC CXR-LT datasets. These results show the value and validity of our ProtoRadNet model.