A new AI model lets two breast-ultrasound algorithms compare notes while they work, and the extra chatter actually helps.
Researchers built a system that handles two jobs at once: finding the edges of a lesion in an ultrasound image and classifying whether it looks benign or malignant. Most prior models split those jobs after a shared starting point, so by the time each branch reaches its final layers they no longer exchange information. The new design adds a Task Interaction Module at four stages of decoding that feeds boundary context into the classification branch and lets class predictions reshape the segmentation channels. An Adaptive Interaction Weighting unit then decides, image by image, how much of that borrowed information to actually use.
On two public datasets, BUSI and BUSI-WHU, the approach beat comparable multi-task, transformer, and decoder-interaction baselines, reaching 86.40% IoU and 95.00% accuracy on the larger set. The ablation numbers are the more telling part: multi-scale context and cross-task sharing each add a little alone, but together they contribute 6.76 points of IoU versus 4.00 when applied separately, and adding the adaptive weighting step pushes classification AUC from 94.41% to 97.31%. That gap suggests the benefit comes from letting the two tasks adjust to each other per image, not just from throwing more computation at the problem.
It is a tidy result on two long-standing benchmark sets, not a clinical trial, so treat the accuracy numbers as a ceiling for a lab setting rather than a verdict on hospital ultrasound machines.