D-MASTER: Mask Annealed Transformer for Unsupervised Domain Adaptation in Breast Cancer Detection from Mammograms

TL;DR

This work tackles unsupervised domain adaptation for breast cancer detection from mammograms, addressing domain shifts and the challenge of small lesion regions. It introduces D-MASTER, a transformer-based Domain-invariant Mask Annealed Student-Teacher Autoencoder that leverages a Deformable DETR backbone, adversarial domain discriminators, and selective retraining to align source and target domains. A novel Mask Annealed Autoencoder (MAE) branch masks multi-scale feature maps with a curriculum and reconstructs the masked last DefDETR layer via a shared mask query, promoting robust target representations, while adaptive confidence refinement gradually shifts pseudo-label filtering from soft to hard using a dynamic weight $\delta$ tied to training progress. Evaluations across INBreast, DDSM, RSNA-BSD1K, and in-house data show substantial gains in sensitivity at $0.3$ FPI over state-of-the-art UDA methods (e.g., $+9\%$ and $+13\%$ on INBreast and DDSM; $+11\%$ and $+17\%$ on in-house and RSNA-BSD1K) and include the RSNA-BSD1K bounding-box annotations, indicating strong potential for practical cross-domain BCDM deployment.

Abstract

We focus on the problem of Unsupervised Domain Adaptation (\uda) for breast cancer detection from mammograms (BCDM) problem. Recent advancements have shown that masked image modeling serves as a robust pretext task for UDA. However, when applied to cross-domain BCDM, these techniques struggle with breast abnormalities such as masses, asymmetries, and micro-calcifications, in part due to the typically much smaller size of region of interest in comparison to natural images. This often results in more false positives per image (FPI) and significant noise in pseudo-labels typically used to bootstrap such techniques. Recognizing these challenges, we introduce a transformer-based Domain-invariant Mask Annealed Student Teacher autoencoder (D-MASTER) framework. D-MASTER adaptively masks and reconstructs multi-scale feature maps, enhancing the model's ability to capture reliable target domain features. D-MASTER also includes adaptive confidence refinement to filter pseudo-labels, ensuring only high-quality detections are considered. We also provide a bounding box annotated subset of 1000 mammograms from the RSNA Breast Screening Dataset (referred to as RSNA-BSD1K) to support further research in BCDM. We evaluate D-MASTER on multiple BCDM datasets acquired from diverse domains. Experimental results show a significant improvement of 9% and 13% in sensitivity at 0.3 FPI over state-of-the-art UDA techniques on publicly available benchmark INBreast and DDSM datasets respectively. We also report an improvement of 11% and 17% on In-house and RSNA-BSD1K datasets respectively. The source code, pre-trained D-MASTER model, along with RSNA-BSD1K dataset annotations is available at https://dmaster-iitd.github.io/webpage.

Figures (5)

• Figure 1: (a) and (b) depict false positive predictions by current teacher student models in cross-domain BCDM. Red boxes indicate ground truth, yellow boxes show MRT zhao2023masked predictions, and green boxes indicate predictions from D-MASTER. As shown in (c), our approach effectively mitigates the domain gap and makes accurate predictions.
• Figure 2: Architecture of proposed Domain-invariant Mask Annealed Student Teacher Autoencoder (D-MASTER) framework
• Figure 3: Mask Annealing updating process
• Figure 4: FROC curves
• Figure S1: Qualitative result comparison on in-house, DDSM, and RSNA-BSD1K datasets. Red boxes show the ground truth, and blue boxes show the predictions.