Medical Image Quality Assessment based on Probability of Necessity and Sufficiency

Boyu Chen,Ameenat L. Solebo,Weiye Bao,Paul Taylor

2024-10-11

Abstract:Medical image quality assessment (MIQA) is essential for reliable medical image analysis. While deep learning has shown promise in this field, current models could be misled by spurious correlations learned from data and struggle with out-of-distribution (OOD) scenarios. To that end, we propose an MIQA framework based on a concept from causal inference: Probability of Necessity and Sufficiency (PNS). PNS measures how likely a set of features is to be both necessary (always present for an outcome) and sufficient (capable of guaranteeing an outcome) for a particular result. Our approach leverages this concept by learning hidden features from medical images with high PNS values for quality prediction. This encourages models to capture more essential predictive information, enhancing their robustness to OOD scenarios. We evaluate our framework on an Anterior Segment Optical Coherence Tomography (AS-OCT) dataset for the MIQA task and experimental results demonstrate the effectiveness of our framework.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The main problems that this paper attempts to solve are two challenges of current deep - learning models in medical image quality assessment (MIQA):

1. **Capturing irrelevant features**: Deep - learning models may learn features irrelevant to image quality from the training data, which may affect the performance of MIQA.
2. **Insufficient ability to handle out - of - distribution (OOD) scenarios**: Deep - learning models may fail and be unable to accurately assess image quality when facing images with a different distribution from the training data.

To solve these problems, the authors propose an MIQA framework based on the causal - inference concept - **Probability of Necessity and Sufficiency (PNS)**. Specifically, this framework predicts image quality by learning hidden features with high PNS values, thereby encouraging the model to capture more essential prediction information and improving its robustness in OOD scenarios.

### Core content of the paper

1. **Introducing the PNS concept**:
- PNS measures whether a set of features are both **necessary** (always present to ensure the result) and **sufficient** (able to guarantee the result). The formula is as follows:
\[
\text{PNS}(Y, H)=\frac{P(Y_{\text{do}(H = h)}=y|H=\bar{h}, Y\neq y)P(H = \bar{h}, Y\neq y)+P(Y_{\text{do}(H=\bar{h})}\neq y|H = h, Y = y)P(H = h, Y = y)
\]
where \(Y\) is the result, \(H\) is the hidden feature, \(h\) and \(\bar{h}\) are the realization of \(H\) and its complement respectively, and \(\text{do}(·)\) represents the intervention operation on the variable \(H\).

2. **Proposing the MIQA - PNS framework**:
- This framework consists of three main parts: a feature extractor, a predictor, and a complementary feature extractor.
- The objective function \(L_{\text{task}}\) includes three loss terms:
\[
L_{\text{task}} := L_{\text{pred}}+L_{\text{compl}}+L_{\text{mono}}
\]
- **Prediction loss \(L_{\text{pred}}\)**: Evaluates the performance of the model in the MIQA task.
- **Complementary loss \(L_{\text{compl}}\)**: Used to evaluate PNS.
- **Monotonicity - constraint loss \(L_{\text{mono}}\)**: Ensures that the features satisfy the monotonicity condition.

3. **Experimental verification**:
- Experiments were carried out using the Anterior Segment Optical Coherence Tomography (AS - OCT) dataset, demonstrating the effectiveness of the MIQA - PNS framework in improving MIQA performance and OOD generalization ability.

### Summary

This paper aims to develop a new MIQA framework by introducing the PNS concept to overcome the limitations of existing deep - learning models in medical image quality assessment. The experimental results show that the MIQA - PNS framework can significantly improve the prediction performance and robustness of the model.