Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

This actually looks like that: Proto-BagNets for local and global interpretability-by-design

Kerol Djoumessi, Bubacarr Bah, Laura Kühlewein, Philipp Berens, Lisa Koch

TL;DR

The Proto-BagNet is introduced, an interpretable-by-design prototype-based model that combines the advantages of bag-of-local feature models and prototype learning to provide meaningful, coherent, and relevant prototypical parts needed for accurate and interpretable image classification tasks.

Abstract

Interpretability is a key requirement for the use of machine learning models in high-stakes applications, including medical diagnosis. Explaining black-box models mostly relies on post-hoc methods that do not faithfully reflect the model's behavior. As a remedy, prototype-based networks have been proposed, but their interpretability is limited as they have been shown to provide coarse, unreliable, and imprecise explanations. In this work, we introduce Proto-BagNets, an interpretable-by-design prototype-based model that combines the advantages of bag-of-local feature models and prototype learning to provide meaningful, coherent, and relevant prototypical parts needed for accurate and interpretable image classification tasks. We evaluated the Proto-BagNet for drusen detection on publicly available retinal OCT data. The Proto-BagNet performed comparably to the state-of-the-art interpretable and non-interpretable models while providing faithful, accurate, and clinically meaningful local and global explanations. The code is available at https://github.com/kdjoumessi/Proto-BagNets.

This actually looks like that: Proto-BagNets for local and global interpretability-by-design

TL;DR

The Proto-BagNet is introduced, an interpretable-by-design prototype-based model that combines the advantages of bag-of-local feature models and prototype learning to provide meaningful, coherent, and relevant prototypical parts needed for accurate and interpretable image classification tasks.

Abstract

Interpretability is a key requirement for the use of machine learning models in high-stakes applications, including medical diagnosis. Explaining black-box models mostly relies on post-hoc methods that do not faithfully reflect the model's behavior. As a remedy, prototype-based networks have been proposed, but their interpretability is limited as they have been shown to provide coarse, unreliable, and imprecise explanations. In this work, we introduce Proto-BagNets, an interpretable-by-design prototype-based model that combines the advantages of bag-of-local feature models and prototype learning to provide meaningful, coherent, and relevant prototypical parts needed for accurate and interpretable image classification tasks. We evaluated the Proto-BagNet for drusen detection on publicly available retinal OCT data. The Proto-BagNet performed comparably to the state-of-the-art interpretable and non-interpretable models while providing faithful, accurate, and clinically meaningful local and global explanations. The code is available at https://github.com/kdjoumessi/Proto-BagNets.
Paper Structure (14 sections, 1 equation, 8 figures, 2 tables)

This paper contains 14 sections, 1 equation, 8 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Developing a faithful prototype-based network
  3. Baseline ProtoPNet model
  4. Enhancing interpretability with the BagNet backbone
  5. Integrating recent advances in training prototype-based models
  6. Results
  7. Dataset
  8. Proto-BagNet yields good accuracy on drusen detection
  9. Proto-BagNet provides understandable localized explanations
  10. Proto-BagNet learns meaningful and relevant prototypes
  11. Proto-BagNet relies on a faithful decision-making process
  12. Discussion and Conclusion
  13. Acknowledgments.
  14. Disclosure of Interests.

Figures (8)

  • Figure 1: Architecture of the Proto-BagNet. (a) Example OCT B-scan image. The red patch illustrates the small receptive field of (b) the BagNet backbone. (c) Feature map and (d) Prototype layer with $m$ prototypes per class. (e) Resulting similarity maps from each prototype to the input. (f) The soft aggregation layer aggregates the average top-k scores from each similarity map into their allocated categories for classification.
  • Figure 1: Examples of some learned prototypes without adding the dissimilarity loss to prevent the model from learning redundant prototypes. (a,b) Prototypes 2,3, and 4 are duplicated. (c,d) Prototypes 2 and 4 are duplicated.
  • Figure 2: Example explanations of ProtoPNet and Proto-BagNet. (a,d) show two learned prototypes with the highest classification weights. Proto-BagNet's prototypes were magnified for visualization only. (b,e) show bounding boxes around regions of highest activation using the visualization technique provided by ProtoPNet (green) and the model's receptive field (yellow). (c,f) show prototypes activations of the query image.
  • Figure 2: Annotated training images from which the disease prototypes were extracted. The green boxes indicate the region where the learned prototypes were extracted, which are enlarged at the bottom. The red markers denote the reference annotations of drusen lesions. The number at the top indicates the prototype ID. For prototype 4, the bounding box is slightly above the lesion, probably due to a mistake when clicking on the lesion.
  • Figure 3: We show (a) the five learned disease prototypes and (b) suspicious regions (green boxes, enlarged below) extracted from each prototype similarity map on an example image. Drusen (annotated with red markers) are detected with high precision.
  • ...and 3 more figures