UniAR leverages a large multimodal model to generate hierarchical diagnostic descriptions at the word, phrase, and sentence levels, compensating for the lack of paired clinical reports. A Mixture‑of‑Experts based Multi‑Scale Alignment Module then dynamically matches vector‑quantized visual prototypes with semantic representations of corresponding granularity, aligning visual evidence with generated semantics.
Extensive experiments on four benchmarks—covering brain MRI and facial expression scenarios—show that UniAR achieves an average accuracy of 75.9% on MRI benchmarks and 91.6% on facial benchmarks, improving over the strongest baselines by 1.5 and 1.2 percentage points respectively. The framework thus offers a robust and interpretable solution for ASD screening under semantic scarcity.
Review