[Paper Note] Interpretable Mammographic Image Classification using Case-Based Reasoning and Deep Learning

Before reading

1. This paper is an extention of ProtoPNet [2]. They applied small changes to fit the model on mammographic image classification.

Overview of interpretable AI algorithm for breast lesions (IAIA-BL)

The same concept of figure is used as the one in ProtoPNet:

To adapt to a my version:

Changes compared to original ProtoPNet

1. Pick up the top-k to do average pooling instead of max pooling for similarity score.

2. Introduce fine annotation loss.
   

3. Add one more fully connected layer before the prediction to calculate the score for each mass-margin type.

Result

Mass-margin

Malignancy

Mathematically

Reference

[1] Interpretable Mammographic Image Classification using Case-Based Reasoning and Deep Learning (https://arxiv.org/abs/2107.05605)

[2] This Looks Like That: Deep Learning for Interpretable Image Recognition (https://arxiv.org/abs/1806.10574)