BCMDA: Bidirectional Correlation Maps Domain Adaptation for Mixed Domain Semi-Supervised Medical Image Segmentation
New method tackles domain shift in medical imaging, showing excellent performance with very few labeled samples.
A team of researchers has introduced BCMDA (Bidirectional Correlation Maps Domain Adaptation), a new framework designed to solve a critical problem in medical AI: training accurate segmentation models when labeled data is scarce and comes from varied sources. The core challenge, known as 'domain shift,' occurs when a model trained on data from one hospital or scanner type fails on data from another. BCMDA tackles this by creating a 'virtual domain' that bridges the gap between labeled and unlabeled data. It uses bidirectional correlation maps to synthesize new images and employs strategies like progressive dynamic MixUp and dual bidirectional CutMix to facilitate smoother knowledge transfer.
On the other hand, to prevent the model from reinforcing its own incorrect predictions (confirmation bias), BCMDA incorporates Prototypical Alignment and Pseudo Label Correction (PAPLC). This technique uses learnable prototype classifiers to align features between the virtual and real domains, creating more compact and reliable representations. It then corrects the model's own pseudo-labels based on these prototypes, leading to more trustworthy training signals from unlabeled data. Empirical results on three public multi-domain datasets demonstrate the framework's superiority, showing it maintains high performance even when very few labeled samples are available, which is often the reality in clinical settings.
- Uses 'Virtual Domain Bridging' with bidirectional correlation maps to synthesize data and align distributions between labeled and unlabeled sets.
- Employs 'Prototypical Alignment' to correct pseudo-labels and combat confirmation bias, creating more reliable training signals.
- Demonstrated excellent performance on three public datasets, specifically excelling in scenarios with extremely limited labeled data.
Why It Matters
This reduces the need for expensive, expert-annotated medical data, accelerating the development of robust AI diagnostic tools for diverse clinical environments.