Multi-Kernel Gated Decoder Adapters for Robust Multi-Task Thyroid Ultrasound under Cross-Center Shift

A research team from the University of British Columbia has published a breakthrough paper on arXiv addressing a fundamental challenge in medical AI: maintaining diagnostic accuracy when AI models trained at one hospital are deployed at another. Their work, "Multi-Kernel Gated Decoder Adapters for Robust Multi-Task Thyroid Ultrasound under Cross-Center Shift," identifies that conventional multi-task learning approaches using a single shared backbone suffer from "negative transfer"—where learning one task (like nodule segmentation) interferes with another (malignancy assessment) when ultrasound data comes from different machines or protocols.

The researchers discovered a consistent pattern: Vision Transformers (ViTs) like MedSAM better preserve geometric priors needed for segmentation, while CNNs like ResNet34 more reliably maintain texture cues crucial for malignancy discrimination under strong domain shifts. To leverage both strengths, they designed lightweight decoder-side adapters called Multi-Kernel Gated Adapters (MKGA) and a residual variant (ResMKGA). These adapters refine multi-scale features using complementary receptive fields and apply semantic, context-conditioned gating to suppress artifact-prone content before fusion.

Across two thyroid ultrasound benchmarks, the proposed adapters demonstrated significant improvements in cross-center robustness. They strengthened out-of-domain segmentation performance and, in the CNN setting, delivered clear gains in clinical TI-RADS diagnostic accuracy compared to standard multi-task baselines. The approach represents a practical solution to domain adaptation without requiring complete model retraining, using only lightweight adapter modules that can be fine-tuned efficiently. Code and models will be released publicly, potentially accelerating adoption of reliable AI-assisted thyroid cancer screening across diverse healthcare institutions.