Unsupervised anomaly localization in high-resolution breast scans using deep pluralistic image completion

TL;DR

This paper tackles unsupervised anomaly localization in ultra-high-resolution DBT by reframing detection as an image-completion problem. It introduces PICARD, which samples multiple normal completions for each patch via spatial dropout and uses the minimum completion distance (MCD) in a learned feature space to score anomalies, with theoretical convergence guarantees as the number of samples grows. Empirically, PICARD achieves state-of-the-art pixel-level localization on the BCS-DBT dataset (e.g., 0.875 pixel AUC with M=10, image-space) and outperforms industrial benchmarks like CutPaste and PatchSVDD, while offering substantial speed advantages over transformer-based pluralistic backbones. The work provides a principled, scalable framework for high-resolution medical anomaly detection and highlights the value of completion-based strategies in settings with scarce abnormal data. The accompanying codebase enables reproducibility and potential extension to other medical imaging modalities.

Abstract

Automated tumor detection in Digital Breast Tomosynthesis (DBT) is a difficult task due to natural tumor rarity, breast tissue variability, and high resolution. Given the scarcity of abnormal images and the abundance of normal images for this problem, an anomaly detection/localization approach could be well-suited. However, most anomaly localization research in machine learning focuses on non-medical datasets, and we find that these methods fall short when adapted to medical imaging datasets. The problem is alleviated when we solve the task from the image completion perspective, in which the presence of anomalies can be indicated by a discrepancy between the original appearance and its auto-completion conditioned on the surroundings. However, there are often many valid normal completions given the same surroundings, especially in the DBT dataset, making this evaluation criterion less precise. To address such an issue, we consider pluralistic image completion by exploring the distribution of possible completions instead of generating fixed predictions. This is achieved through our novel application of spatial dropout on the completion network during inference time only, which requires no additional training cost and is effective at generating diverse completions. We further propose minimum completion distance (MCD), a new metric for detecting anomalies, thanks to these stochastic completions. We provide theoretical as well as empirical support for the superiority over existing methods of using the proposed method for anomaly localization. On the DBT dataset, our model outperforms other state-of-the-art methods by at least 10\% AUROC for pixel-level detection.

Figures (8)

• Figure 1: The outer loop of our proposed anomaly localization method, PICARD (Algorithm \ref{['alg:mindist']}), at the slice level. See Figure \ref{['fig:model_patch']} for the inner loop at the patch level.
• Figure 2: The inner loop of our proposed anomaly localization method, at the patch level. See Figure \ref{['fig:model_slice']} for the outer loop at the slice level.
• Figure 3: Examples of pluralistic normal completions of a normal DBT patch (top block) and an anomalous patch (bottom block) using our method (Section \ref{['sec:theory:dropout_inpaint']}) and HFPIC wan2021high. Left column: input image, masked and unmasked; center column: completions with our method; right column: completions with HFPIC. Image contrast modified to improve visibility.
• Figure 4: Histogram (left) and associated AUC (area under the receiver operating characteristic curve, right) of a particular test DBT cancer slice (the top left row of Fig. \ref{['fig:comparemodels']}), for the normalized distributions of MCD anomaly metric scores for normal pixels (blue) and anomalous pixels (red).
• Figure 5: Qualitative tumor localization performance for our method (PICARD) compared to several state-of-the-art methods.For each example test image, we show the performance (from left to right) of (1) our method, PICARD; (2) PICARD with the deterministic, single-completion case; (3) CutPaste li2021cutpaste; and (4) PatchSVDD patchsvdd. The two examples on the bottom row demonstrate performance on cases with dense breast tissue. Refer to Table \ref{['tab:modelcompare']} for corresponding quantitative results on the entire test set. This figure is best viewed in color.