Decoding Decision Reasoning: A Counterfactual-Powered Model for Knowledge Discovery

TL;DR

This work proposes an explainable model that is equipped with both decision reasoning and feature identification capabilities, and validated in the demanding realm of medical prognosis task, demonstrating its efficacy and potential in enhancing the reliability of AI in healthcare and in discovering new knowledge in diseases where prognostic understanding is limited.

Abstract

In medical imaging, particularly in early disease detection and prognosis tasks, discerning the rationale behind an AI model's predictions is crucial for evaluating the reliability of its decisions. Conventional explanation methods face challenges in identifying discernible decisive features in medical image classifications, where discriminative features are subtle or not immediately apparent. To bridge this gap, we propose an explainable model that is equipped with both decision reasoning and feature identification capabilities. Our approach not only detects influential image patterns but also uncovers the decisive features that drive the model's final predictions. By implementing our method, we can efficiently identify and visualise class-specific features leveraged by the data-driven model, providing insights into the decision-making processes of deep learning models. We validated our model in the demanding realm of medical prognosis task, demonstrating its efficacy and potential in enhancing the reliability of AI in healthcare and in discovering new knowledge in diseases where prognostic understanding is limited.

Figures (3)

• Figure 1: (A) The decision-decoding workflow of PrognosisEx; (B) The training process of PrognosisEx, which includes the development of a general autoencoder that compresses slices into a low-dimensional vector, followed by the training of a two-layer classifier model.
• Figure 2: Identification of class-specific attributes in $\mathbb{V}_c$ space for our binary prognostic task to predict the mortality of COVID-19 patients in 10 days, where class 1 represents the label of death and class 0 represents the label of survival. In this case, $s_{2,1}$ and $s_{2,2}$ in $\mathbb{V}_1$ space are class 1-specific features, which are regarded as most decisive features to decision 1. Similarly, $s_{2,2}$, $s_{3,1}$, and $s_{3,2}$ in $\mathbb{V}_0$ space are class 0-specific features, considered the most decisive features to decision 0.
• Figure 3: Visualisation of identified 'class-specific' features in the COVID-19 prognostic task using PrognosisEx. Figure (a) and (b) depict counterfactuals demonstrating the enhanced and mitigated contribution of 'Class 1-Specific' Features (features contributing to the death decision); Figure (c) and (d) depict counterfactuals demonstrating the enhanced and mitigated influence of 'Class 0-Specific' features (features contributing to survival decision). In each image, the 1st column shows the image reconstructed from the unmanipulated feature, with the following two columns the images reconstructed from the manipulated features.