Exploring Pre-trained models on Ultrasound Modeling for Mice Autism Detection with Uniform Filter Bank and Attentive Scoring
Genetically engineered mice, whose behaviors resemble those of individuals with Autism Spectrum Disorder (ASD), serve as valuable models for studying ASD through ultrasound vocalization (USV) analysis. In this paper, we investigate the effectiveness of pre-trained models in learning features of the USVs by fine-tuning. To bridge the gap between the pre-trained model and the inductive bias of the ultrasonic signal, we design a uniformly-spaced filter bank to reduce the dimension in the frequency domain. The extracted filter-bank energies of the ultrasonic spectrogram form a pseudo-spectrogram for pre-trained models. In the back-end, we employ an attentive frame-wise scoring method for classification, resulting in a comprehensive judgment. Experimental results demonstrate the effectiveness of our approach, achieving a segment-level Unweighted Average Recall (UAR) of 0.729 and a subject-level UAR of 0.882 on the validation set provided by the MADUV 2025 Challenge.
