Distilling BlackBox to Interpretable Models for Efficient Transfer Learning

Shantanu Ghosh, Ke Yu, Kayhan Batmanghelich

Introduction

Deep learning has demonstrated remarkable success in medical imaging, particularly in interpreting Chest X-rays (CXR). However, these models face two significant hurdles: a lack of interpretability and "brittleness" under distribution shifts. When a model trained at one institution is deployed at another, differences in hardware, protocols, and demographics often lead to performance degradation. Addressing this typically requires fine-tuning the entire network on a large, expertly labeled dataset from the new domain—a process that is both computationally expensive and data-intensive.

This paper introduces MoIE-CXR (Mixture of Interpretable Experts), a framework designed to bridge the gap between high-performing black-box models and the need for interpretable, adaptable clinical tools. By distilling complex neural networks into a series of shallow, logic-based experts, the authors achieve state-of-the-art performance with a fraction of the computational overhead.

The MoIE-CXR Framework

The core of the proposed approach is a "Route, Interpret, Repeat" cycle that iteratively extracts knowledge from a pre-trained black-box model.

1. Iterative Distillation

Instead of replacing the black-box model with a single interpretable one, MoIE-CXR creates a mixture of experts. Each expert is a shallow model that utilizes First-Order Logic (FOL) rules based on 107 clinical concepts (e.g., "pleural effusion," "enlarged heart") extracted from radiology reports using the RadGraph tool.

2. The Selector Mechanism

A "selector" network accompanies each expert. For any given input, the selector determines whether the current expert can provide a confident and accurate explanation. If the expert is deemed sufficient, the prediction is made; if not, the sample is passed to a "residual" (the remaining portion of the black-box model) to be handled in the next iteration. This ensures that "easy" cases are handled by simple logic, while "hard" cases are reserved for more complex processing.

3. Stratified Coverage

To prevent the model from ignoring rare diseases—a common issue in medical datasets—the authors implement a stratified coverage approach. This ensures that the experts are trained to handle a diverse range of clinical conditions, maintaining high sensitivity even for infrequent classes.

Efficient Transfer Learning

The most significant advantage of MoIE-CXR lies in its ability to adapt to new domains. Traditional fine-tuning requires updating millions of parameters. In contrast, MoIE-CXR focuses on the premise that while image characteristics change across hospitals, the underlying clinical logic remains constant.

When moving to a new target domain:

  1. Pseudo-labeling: A concept detector is trained using semi-supervised techniques to identify the 107 clinical concepts in the new images.
  2. Targeted Fine-tuning: Only the small, interpretable experts and their selectors are fine-tuned. Because these models are shallow and logic-based, they require significantly less data and compute to reach peak performance.

Experimental Evaluation

The researchers validated MoIE-CXR using two massive, real-world datasets:

  • Source Domain: MIMIC-CXR (over 220,000 images).
  • Target Domain: Stanford CheXpert.
  • Concepts: 107 anatomical and observation concepts were derived using the RadGraph tool.

Results

The experiments yielded three primary findings:

1. Performance Parity and Improvement

MoIE-CXR does not sacrifice accuracy for interpretability. For conditions like pleural effusion, the model achieved an AUROC of 0.93, surpassing the original black-box model (0.92) and significantly outperforming traditional Concept Bottleneck Models (CBMs).

2. Extreme Computational Efficiency

The computational savings are unprecedented. When fine-tuning on 2.5% of the target data, a standard black-box model required 776 Tera-FLOPs. MoIE-CXR achieved superior results using only 0.0065 Tera-FLOPs—a 120,000x reduction in computational cost.

3. Data Efficiency

MoIE-CXR proved highly robust in low-data regimes. With only 2.5% of the target domain's labeled data, the fine-tuned MoIE-CXR outperformed a fully fine-tuned black-box model by an average of 5%.

Discussion and Limitations

While MoIE-CXR represents a major step forward, the authors note several limitations. The model relies on correlations between clinical concepts and diagnoses; these relationships are not necessarily causal. Furthermore, the system's effectiveness is bound by the quality and comprehensiveness of the initial concept list. If a critical diagnostic feature is missing from the 107 predefined concepts, the interpretable experts may fail to capture the full clinical picture.

Future research will focus on incorporating causal discovery to ensure the logic used by the experts aligns more closely with medical causality, further enhancing the reliability and trustworthiness of AI in clinical settings.

Summary updated

BibTeX
@inproceedings{ghosh2023distilling,
  author = {Shantanu Ghosh and Ke Yu and Kayhan Batmanghelich},
  title = {{Distilling BlackBox to Interpretable Models for Efficient Transfer Learning}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  url = {https://openreview.net/forum?id=e0bEejCTLS},
}
Figure