Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PRECISe : Prototype-Reservation for Explainable Classification under Imbalanced and Scarce-Data Settings

Vaibhav Ganatra, Drishti Goel

TL;DR

The paper addresses medical image classification under severe data scarcity and class imbalance while demanding faithful explanations for clinical trust. It introduces PRECISe, an explainable-by-design architecture that combines an autoencoder, a prototype-metric layer, and class-reserved prototypes to generate faithful, human-interpretable explanations via distances to prototypes in latent space. Across two imbalanced datasets (Pneumonia chest X-ray and BUSI ultrasound), PRECISe achieves state-of-the-art data-efficient performance (e.g., 92.04% and 88.75% accuracy) and maintains high minority-class accuracy with very small training sets, while providing tangible prototype-based explanations. The work demonstrates that prototype-based explanations can be produced without post-hoc methods and highlights potential clinical impact in settings with limited labeled data.

Abstract

Deep learning models used for medical image classification tasks are often constrained by the limited amount of training data along with severe class imbalance. Despite these problems, models should be explainable to enable human trust in the models' decisions to ensure wider adoption in high-risk situations. In this paper, we propose PRECISe, an explainable-by-design model meticulously constructed to concurrently address all three challenges. Evaluation on 2 imbalanced medical image datasets reveals that PRECISe outperforms the current state-of-the-art methods on data efficient generalization to minority classes, achieving an accuracy of ~87% in detecting pneumonia in chest x-rays upon training on <60 images only. Additionally, a case study is presented to highlight the model's ability to produce easily interpretable predictions, reinforcing its practical utility and reliability for medical imaging tasks.

PRECISe : Prototype-Reservation for Explainable Classification under Imbalanced and Scarce-Data Settings

TL;DR

The paper addresses medical image classification under severe data scarcity and class imbalance while demanding faithful explanations for clinical trust. It introduces PRECISe, an explainable-by-design architecture that combines an autoencoder, a prototype-metric layer, and class-reserved prototypes to generate faithful, human-interpretable explanations via distances to prototypes in latent space. Across two imbalanced datasets (Pneumonia chest X-ray and BUSI ultrasound), PRECISe achieves state-of-the-art data-efficient performance (e.g., 92.04% and 88.75% accuracy) and maintains high minority-class accuracy with very small training sets, while providing tangible prototype-based explanations. The work demonstrates that prototype-based explanations can be produced without post-hoc methods and highlights potential clinical impact in settings with limited labeled data.

Abstract

Deep learning models used for medical image classification tasks are often constrained by the limited amount of training data along with severe class imbalance. Despite these problems, models should be explainable to enable human trust in the models' decisions to ensure wider adoption in high-risk situations. In this paper, we propose PRECISe, an explainable-by-design model meticulously constructed to concurrently address all three challenges. Evaluation on 2 imbalanced medical image datasets reveals that PRECISe outperforms the current state-of-the-art methods on data efficient generalization to minority classes, achieving an accuracy of ~87% in detecting pneumonia in chest x-rays upon training on <60 images only. Additionally, a case study is presented to highlight the model's ability to produce easily interpretable predictions, reinforcing its practical utility and reliability for medical imaging tasks.
Paper Structure (13 sections, 4 equations, 7 figures, 5 tables)

This paper contains 13 sections, 4 equations, 7 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Network Architecture
  5. Network Optimization - Prototype Reservation
  6. Experiments and Results
  7. Datasets
  8. Baselines and Evaluation
  9. Implementation
  10. Results and Discussion
  11. Explainability Case Study
  12. Limitations
  13. Conclusion

Figures (7)

  • Figure 1: Class-wise data distribution from the RetinaMNIST dataset. Severity-1 Diabetic Retinopathy (DR) makes up 45% of the dataset, whereas only $\sim$6% (and only 66 images) of the data belongs to Severity-5 DR
  • Figure 2: Overview of Prototype-Reservation
  • Figure 3: Performance on subsets of varying sizes of the Pneumonia dataset. PRECISe (ours) shows excellent performance retention with reducing training set sizes
  • Figure 4: Performance on subsets of varying sizes of the BUSI dataset. PRECISe (ours) shows the best performance at all training set sizes.
  • Figure 5: Classwise performance on the Pneumonia dataset. PRECISe (ours) displays superior performance on the minority classes.
  • ...and 2 more figures