Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SHAP-Integrated Convolutional Diagnostic Networks for Feature-Selective Medical Analysis

Yan Hu, Ahmad Chaddad

TL;DR

This work tackles the challenge of developing accurate medical image classifiers under privacy-driven data limits. It introduces SICDN, a SHAP-integrated CNN framework that performs feature selection by weighting the fully connected layer with a SHAP-derived matrix $S^{*}$ and updates using Adam, augmented by a historical weighted moving average controlled by $\lambda$. Across pneumonia and breast cancer datasets, SICDN achieves state-of-the-art performance on limited data, notably attaining a subset pneumonia AUC near 99.66% and high accuracy and F1, while remaining competitive on full datasets. The approach enhances interpretability and efficiency, and the authors provide code to enable reproducibility and potential integration with privacy-preserving paradigms like Federated Learning in clinical settings.

Abstract

This study introduces the SHAP-integrated convolutional diagnostic network (SICDN), an interpretable feature selection method designed for limited datasets, to address the challenge posed by data privacy regulations that restrict access to medical datasets. The SICDN model was tested on classification tasks using pneumonia and breast cancer datasets, demonstrating over 97% accuracy and surpassing four popular CNN models. We also integrated a historical weighted moving average technique to enhance feature selection. The SICDN shows potential in medical image prediction, with the code available on https://github.com/AIPMLab/SICDN.

SHAP-Integrated Convolutional Diagnostic Networks for Feature-Selective Medical Analysis

TL;DR

This work tackles the challenge of developing accurate medical image classifiers under privacy-driven data limits. It introduces SICDN, a SHAP-integrated CNN framework that performs feature selection by weighting the fully connected layer with a SHAP-derived matrix and updates using Adam, augmented by a historical weighted moving average controlled by . Across pneumonia and breast cancer datasets, SICDN achieves state-of-the-art performance on limited data, notably attaining a subset pneumonia AUC near 99.66% and high accuracy and F1, while remaining competitive on full datasets. The approach enhances interpretability and efficiency, and the authors provide code to enable reproducibility and potential integration with privacy-preserving paradigms like Federated Learning in clinical settings.

Abstract

This study introduces the SHAP-integrated convolutional diagnostic network (SICDN), an interpretable feature selection method designed for limited datasets, to address the challenge posed by data privacy regulations that restrict access to medical datasets. The SICDN model was tested on classification tasks using pneumonia and breast cancer datasets, demonstrating over 97% accuracy and surpassing four popular CNN models. We also integrated a historical weighted moving average technique to enhance feature selection. The SICDN shows potential in medical image prediction, with the code available on https://github.com/AIPMLab/SICDN.

Paper Structure

This paper contains 10 sections, 9 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Feature extraction and selection
  4. Weights update
  5. Historical weighted moving average
  6. Experiments
  7. Dataset
  8. Experimental setup
  9. Results
  10. Conclusion

Figures (2)

  • Figure 1: Flowchart of the training for SICDN. (1) Transform the medical images and extract features using the trained model, (2) Flatten the features and compute the Shapley values, (3) Normalize the absoluted Shapley value, (4) Use the normalized Shapley values as an importance matrix to multiply with the weights matrix of the fully connected layer and update the weights.
  • Figure 2: The ROC curves and top AUC scores of the models.