Abstract
The auditory brainstem response (ABR) is an important objective tool for hearing assessment, but its manual interpretation remains labor-intensive and subject to inter-observer variability. Although deep learning approaches applied to the actual ABR voltage measurements have shown promising performance, their clinical use may be constrained by limited access to raw data, as commercial ABR systems often store recordings in vendor-specific formats and require proprietary software or device-specific export procedures. In contrast, ABR report images are routinely generated in clinical practice and provide a more accessible source for automated analysis. However, existing image-based approaches often struggle to capture long-range waveform dependencies and may be distracted by non-informative backgrounds. To address these challenges, we propose CASTS, an interpretable framework for automated ABR report image classification. CASTS uses a pretrained ViT-B/16 backbone with multi-layer feature fusion to integrate hierarchical waveform-image representations. A classification (CLS)-guided top-k patch selection module is also introduced to identify and aggregate waveform-relevant patch tokens according to their similarity to the CLS token, thereby reducing the influence of background-dominant regions. Evaluated on a public dataset containing 186 ABR report images, CASTS achieved a mean classification accuracy of 94.88% and an AUC of 0.976, outperforming representative convolutional and Transformer-based feature extraction baselines. Ablation studies confirmed the effectiveness of the proposed components. Occlusion-based sensitivity analysis further showed that CASTS primarily focused on clinically meaningful waveform structure, especially wave V, rather than non-informative background areas. These results suggest that CASTS provides an effective and interpretable solution for automated ABR report image analysis.

This publication has 27 references indexed in Scilit: