IBO: Inpainting-Based Occlusion to Enhance Explainable Artificial Intelligence Evaluation in Histopathology

TL;DR

This work tackles the reliability gap in evaluating explainable AI for histopathology by addressing OoD artifacts produced by traditional occlusion. It introduces Inpainting-Based Occlusion (IBO), which uses a diffusion model trained on normal tissue and RePaint-based inpainting to replace tumor regions with realistic non-cancer tissue, preserving data distribution. Empirically, IBO achieves substantially better perceptual fidelity ($LPIPS$) and more accurate rankings of XAI heatmaps (via IoU and $AUC$ with a reduced mean absolute rank difference), outperforming existing occlusion strategies. The approach offers a more trustworthy framework for XAI evaluation in histopathology and potentially other imaging domains, with available source code at the project repository.

Abstract

Histopathological image analysis is crucial for accurate cancer diagnosis and treatment planning. While deep learning models, especially convolutional neural networks, have advanced this field, their "black-box" nature raises concerns about interpretability and trustworthiness. Explainable Artificial Intelligence (XAI) techniques aim to address these concerns, but evaluating their effectiveness remains challenging. A significant issue with current occlusion-based XAI methods is that they often generate Out-of-Distribution (OoD) samples, leading to inaccurate evaluations. In this paper, we introduce Inpainting-Based Occlusion (IBO), a novel occlusion strategy that utilizes a Denoising Diffusion Probabilistic Model to inpaint occluded regions in histopathological images. By replacing cancerous areas with realistic, non-cancerous tissue, IBO minimizes OoD artifacts and preserves data integrity. We evaluate our method on the CAMELYON16 dataset through two phases: first, by assessing perceptual similarity using the Learned Perceptual Image Patch Similarity (LPIPS) metric, and second, by quantifying the impact on model predictions through Area Under the Curve (AUC) analysis. Our results demonstrate that IBO significantly improves perceptual fidelity, achieving nearly twice the improvement in LPIPS scores compared to the best existing occlusion strategy. Additionally, IBO increased the precision of XAI performance prediction from 42% to 71% compared to traditional methods. These results demonstrate IBO's potential to provide more reliable evaluations of XAI techniques, benefiting histopathology and other applications. The source code for this study is available at https://github.com/a-fsh-r/IBO.

Figures (6)

• Figure 1: Overview of the methodology used for evaluating XAI methods for breast cancer diagnosis from histopathological images. In Masks, black regions indicate the areas that should be inpainted. These black regions correspond to the tumoric areas identified by the XAI approaches, which are crucial for evaluating the effectiveness of our methodology.
• Figure 2: Comparison of generated heatmap, grayscale version of the heatmap, and levels of importance.
• Figure 3: Side-by-side comparison of the WSI (left) and the mask image (right). In the mask image, the white regions indicate tumoric areas.
• Figure 4: A sequence of images showcasing different CAM-based methods applied to the tumor patch for decision making. Each method provides a unique perspective on the activation areas within the image.
• Figure 5: A sequence showing a tumor patch, its Full-Grad heatmap, and various masks. Line 1 displays the original tumor patch and its corresponding heatmap, with masks highlighting regions of varying importance from high to low. Line 2 illustrates that the black regions in each mask represent areas that have been occluded after each step. This sequence demonstrates the progressive occlusion of tumor regions based on their significance as indicated by the heatmap.