An Interpretable Deep Learning System for Fine-Grained Classification and Longitudinal Tracking of Neonatal Auricular...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Neonatal Medicine

Background:

• Accurate diagnosis of neonatal auricular deformities is crucial for timely non-invasive correction.
• Clinical assessment relies on subjective visual evaluation, leading to missed treatment windows.
• Objective metrics for tracking deformity progression and treatment efficacy are lacking.

Purpose of the Study:

• To develop an interpretable deep learning system for automated screening and classification of neonatal ear deformities.
• To enable objective, quantitative tracking of treatment response.
• To provide a standardized tool for early intervention and precision management.

Main Methods:

• Curated a large-scale dataset (n=4644) for training.
• Employed YOLOv11 for object detection and ConvNeXt-Tiny for classification.
• Integrated supervised contrastive learning for a continuous severity score and Grad-CAM for interpretability.

Main Results:

• Achieved 88.2% accuracy in binary screening (AUC: 0.949) and 87.4% accuracy in multi-class subtyping (macro-AUC: 0.976).
• Demonstrated robust generalization across three independent cohorts.
• Severity scores effectively quantified post-intervention improvements (p=0.0004).

Conclusions:

• The developed AI system offers an objective and standardized approach to neonatal ear deformity diagnosis.
• It facilitates early intervention and precision management by enabling quantitative tracking of therapeutic efficacy.
• Further clinical calibration of the severity score is needed for rare subtypes and broader application.