Ultrasound diagnosis stops chasing average accuracy: multi-query attention keeps heterogeneous evidence, training-time regularization lifts sensitivity to 0.874

Patient-level ultrasound aggregation moves from a single attention bottleneck to multi-query cross-attention, where several learnable queries each act as an evidence extractor and preserve complementary signs across views before classification. Earlier mean pooling, max pooling, or single-query attention compressed heterogeneous evidence through one representation bottleneck, either diluting sparse focal cues or letting a single high-scoring instance dominate, a failure mode called unimodal attention collapse. On the biopsy-proven thyroid dataset ThyroidXL, the four-query design reaches patient-level AUC 0.932, above 0.916 for one query and 0.914 for eight queries; on the breast dataset BUS-BRA it reaches 0.936, above 0.887 and 0.888 for the same two settings.

The training objective adds a differentiable sensitivity-specificity geometric-mean regularizer to cross-entropy, penalizing imbalance between the two error types during training instead of leaving the clinical operating point entirely to post-hoc threshold tuning. Standard cross-entropy minimizes global classification error and, in imbalanced cohorts, tends to favor the majority class, producing an undesirable sensitivity-specificity trade-off. Without the regularizer, ThyroidXL sensitivity is 0.813 and specificity 0.907; with it, sensitivity rises to 0.874 and specificity falls to 0.876, while AUC rises from 0.912 to 0.932.

Instance features come from a DINOv3 foundation model adapted with LoRA, which fine-tunes low-rank parameters in the attention and projection layers rather than freezing or fully fine-tuning the backbone. A frozen backbone reaches AUC 0.877 on ThyroidXL and full fine-tuning 0.926, both below low-rank adaptation. Low-rank adaptation reaches AUC 0.932 and sensitivity 0.874 on ThyroidXL, and AUC 0.936 and sensitivity 0.876 on BUS-BRA; full fine-tuning reaches AUC 0.942 on BUS-BRA, slightly above low-rank adaptation.

Evaluation covers two biopsy-proven ultrasound benchmarks: ThyroidXL with 11,635 images from 4,093 patients and BUS-BRA with 1,875 breast images from 1,064 patients, both reported at the patient level on official splits for AUC, AUPRC, sensitivity, and specificity, with the threshold chosen on an internal validation subset by maximizing the geometric mean and then fixed. Ablations identify four queries, concatenated fusion, and low-rank adaptation as the default configuration, and AUC and geometric mean generally improve as more views are available. It suits researchers and engineers working on medical multi-instance learning and clinical trade-off alignment.

An overly large regularization weight over-constrains optimization, so the paper notes a moderate weight is needed; specificity drops as sensitivity rises, from 0.907 to 0.876 on ThyroidXL and from 0.916 to 0.879 on BUS-BRA. Full fine-tuning reaches AUC 0.942 on BUS-BRA, slightly above the 0.936 of low-rank adaptation, so parameter-efficient adaptation does not lead on every metric.