A research team has built a neural network architecture that lets image segmentation and classification tasks actively teach each other during inference.
Most multi-task medical imaging models share an encoder but keep their decoder branches separate, meaning boundary details discovered by the segmentation branch never reach the classifier, and semantic context from classification never sharpens the segmentation masks. BTI-Net breaks that wall by inserting Task Interaction Modules at all four decoder resolutions, running two parallel pathways that push spatial boundary cues into classification and feed global semantic priors back into segmentation. The trick is bidirectionality: refined features flow progressively from coarse to fine detail across every decoder level, not just at the output stage.
The gains are modest but consistent across three different imaging modalities — ultrasound, dermoscopy, and brain MRI — which matters because most architecture papers cherry-pick a single domain. The team's Uncertainty Proxy Attention mechanism deserves attention: it gates each cross-task signal per input instance using three proxy signals for alignment, scene complexity, and prediction confidence, with no extra inference passes or manual annotations required. That adaptive gating alone accounts for a +2.36 IoU improvement over fixed bidirectional interaction and up to +2.26 percentage points in classification accuracy over the strongest multi-task baseline.
The architecture is a direct challenge to the assumption that sharing encoders is sufficient for multi-task synergy. Whether it holds up outside curated benchmarks — in the messier data that clinical deployments actually see — remains the real test.