AnomExpert: Identifying and Selecting Anatomical Planes for Prenatal Ultrasound Anomaly Diagnosis

Prenatal ultrasound (US) diagnosis is a complex, time-consuming process that requires clinicians to identify specific anatomical planes and determine which ones are relevant for diagnosing congenital anomalies. Existing automated methods often struggle because they either require expensive, fine-grained image-level annotations or fail to explicitly model the relationship between anatomical planes and specific diseases. This paper introduces AnomExpert, a framework designed to perform accurate anomaly classification using only case-level supervision, effectively learning to identify and select diagnostic planes without explicit labels.

AnomExpert addresses the limitations of standard Multi-Instance Learning (MIL) by explicitly modeling two core diagnostic capabilities:

1. Anatomical Plane Identification: The framework uses a set of learnable 'plane prototypes' that organize unordered ultrasound images into latent representations. By using Sinkhorn-balanced normalization, the model clusters images with similar anatomical content, effectively discovering anatomical planes without needing manual annotations.

2. Disease-Aware Sparse Selection: Once images are grouped into plane representations, the model uses learnable 'disease queries' and a bias matrix to estimate the relevance of each plane to a specific anomaly. A sparse selection mechanism then identifies the top-k most diagnostically relevant planes for each case, ensuring the final classification is based on the most informative evidence.

The model is trained end-to-end using only the primary diagnosis label for each case, making it highly efficient for clinical workflows where detailed annotations are unavailable.

Evaluated on a multi-center dataset of 3,654 cases, AnomExpert consistently outperformed nine representative MIL methods. Using a ViT-small backbone, it achieved an accuracy of 86.9% and an F1-score of 84.2%. The study demonstrates that explicitly modeling plane identification and disease-specific selection is superior to traditional aggregation methods. Furthermore, the model maintains high parameter efficiency, with a ViT-tiny version achieving competitive performance with only 5.6M parameters.