Enhancing AI Diagnostics: Autonomous Lesion Masking via Semi-Supervised Deep Learning

TL;DR

The paper tackles data scarcity and annotation burden in breast ultrasound lesion classification by addressing domain shift between public source and private target datasets. It introduces an unsupervised domain adaptation framework that starts with a labeled source dataset to train a segmentation teacher, generates pseudo-masks for unlabeled target data, and uses downstream classification to guide iterative refinement. Techniques include entropy-based pre-processing, hole filling, convex hull post-processing, and a combined training scheme of teacher-student self-training culminating in improved ROI masks and classification performance. The work demonstrates potential to streamline ROI annotation and improve interpretability of AI-based breast lesion diagnosis, with broader applicability to medical imaging domains.

Abstract

This study presents an unsupervised domain adaptation method aimed at autonomously generating image masks outlining regions of interest (ROIs) for differentiating breast lesions in breast ultrasound (US) imaging. Our semi-supervised learning approach utilizes a primitive model trained on a small public breast US dataset with true annotations. This model is then iteratively refined for the domain adaptation task, generating pseudo-masks for our private, unannotated breast US dataset. The dataset, twice the size of the public one, exhibits considerable variability in image acquisition perspectives and demographic representation, posing a domain-shift challenge. Unlike typical domain adversarial training, we employ downstream classification outcomes as a benchmark to guide the updating of pseudo-masks in subsequent iterations. We found the classification precision to be highly correlated with the completeness of the generated ROIs, which promotes the explainability of the deep learning classification model. Preliminary findings demonstrate the efficacy and reliability of this approach in streamlining the ROI annotation process, thereby enhancing the classification and localization of breast lesions for more precise and interpretable diagnoses.

Figures (8)

• Figure 1: Overview of our proposed approach. We train Teacher Model on the labeled dataset $D_L$ and iteratively repeat the following: predict on the unlabeled dataset $D_U$ for pseudo-masks, perform downstream classification, and train Student Model on the combined dataset. Self-training terminates when classification performance is satisfactory.
• Figure 2: Our proposed approach for unsupervised domain adaptation. Given the domain shift between source and target domains, we propose a framework to generate and refine the pseudo-masks for the target domain, which are then used for downstream classification. (The pseudo-mask presented here is obtained after iterations of self-training.)
• Figure 3: Cropping an image with entropy filter. (a) Original Image (b) Entropy filtered image (c) Contours of Entropy filtered image (d) Identification of exam area (e) Final cropped image
• Figure 4: Convex hull operation. (a) Non-convex area (b) Convex hull process (c) Final Convex hull (d) Input Image (e) Pseudo-mask (f) Convex hull of pseudo-mask (g) 1.5x scaled convex hull.
• Figure 5: Intensity distribution of source and pre-processed target images.