FeaInfNet: Diagnosis in Medical Image with Feature-Driven Inference and Visual Explanations (2312.01871v1)

Abstract: Interpretable deep learning models have received widespread attention in the field of image recognition. Due to the unique multi-instance learning of medical images and the difficulty in identifying decision-making regions, many interpretability models that have been proposed still have problems of insufficient accuracy and interpretability in medical image disease diagnosis. To solve these problems, we propose feature-driven inference network (FeaInfNet). Our first key innovation involves proposing a feature-based network reasoning structure, which is applied to FeaInfNet. The network of this structure compares the similarity of each sub-region image patch with the disease templates and normal templates that may appear in the region, and finally combines the comparison of each sub-region to make the final diagnosis. It simulates the diagnosis process of doctors to make the model interpretable in the reasoning process, while avoiding the misleading caused by the participation of normal areas in reasoning. Secondly, we propose local feature masks (LFM) to extract feature vectors in order to provide global information for these vectors, thus enhancing the expressive ability of the FeaInfNet. Finally, we propose adaptive dynamic masks (Adaptive-DM) to interpret feature vectors and prototypes into human-understandable image patches to provide accurate visual interpretation. We conducted qualitative and quantitative experiments on multiple publicly available medical datasets, including RSNA, iChallenge-PM, Covid-19, ChinaCXRSet, and MontgomerySet. The results of our experiments validate that our method achieves state-of-the-art performance in terms of classification accuracy and interpretability compared to baseline methods in medical image diagnosis. Additional ablation studies verify the effectiveness of each of our proposed components.

• The paper introduces a feature-driven reasoning structure that simulates differential diagnosis for medical imaging.
• It employs Local Feature Masks to extract both local and global context, enhancing interpretability.
• Adaptive Dynamic Masks generate refined saliency maps that enable robust lesion localization and diagnostic accuracy.

Understanding FeaInfNet: Interpretable AI for Medical Image Diagnosis

Introduction

Medical image diagnosis is an area where accuracy and interpretability are paramount. AI models, such as Convolutional Neural Networks (CNNs), have been highly successful in image recognition tasks. However, their "black box" nature often makes understanding their decision-making process difficult. This is a significant issue in healthcare, where explanations for diagnoses are essential. Enter FeaInfNet, an interpretable deep learning model aimed at improving both the accuracy and interpretability of medical image diagnosis.

Feature-Based Reasoning Structure

FeaInfNet introduces a feature-based reasoning structure that aims to simulate how doctors diagnose diseases. The structure does this by comparing the similarity of each sub-region image patch with disease and normal templates that might appear in that region, resembling the process of differential diagnosis in human reasoning. This approach avoids the potential pitfalls of prototype-based reasoning, prevalent in previous interpretable models, which might lead to misleading conclusions due to the tendency to take into account disease areas alongside normal areas within medical images.

Local Feature Masks and Adaptive Dynamic Masks

To empower the interpretability of FeaInfNet, the model implements Local Feature Masks (LFM) for extracting feature vectors. This method provides both local and global context to the feature vectors, enhancing the model’s expressive ability.

Moreover, FeaInfNet introduces Adaptive Dynamic Masks (Adaptive-DM) for generating saliency maps. Instead of just upsampling similarity activations as traditional methods do, Adaptive-DM dynamically adjusts the saliency maps to better retain decision-making areas. It weighs the importance between similarity terms and mask terms autonomously, resulting in high-quality visual interpretations.

Empirical Success

The experimental results verify FeaInfNet's effectiveness across multiple public medical datasets. It achieves state-of-the-art performance in terms of classification accuracy and interpretability when compared with baseline methods—including both non-interpretable and prototype-based interpretable networks.

FeaInfNet's unique design enables robust lesion localization. Through its Adaptive-DM process, the network more accurately pinpoints pathological regions used for decision-making. The saliency maps provide a complete and refined visualization of lesions, significantly outperforming those generated by prior techniques.

Conclusion

FeaInfNet represents an important step forward for AI in medical imaging. It provides a model that delivers high recognition performance combined with the crucial element of interpretability. By mimicking the diagnosis process of medical professionals, FeaInfNet assures trust in AI-driven diagnostic systems, offering visual explanations alongside accurate predictions. This bridge between human-understandable reasoning and advanced machine learning could be a harbinger for broader adoption of AI in clinical settings.