AI-Assisted Diagnosis of Neonatal Ear Anomalies Using ResNet50 CBAM and CNN Transformer Models

Early detection of congenital auricular anomaly in infants is essential for early medical intervention, as late diagnosis would have long term psychosocial impacts and facial asymmetry. This study introduces two deep learning based architectures for the automated classification of auricular anomalies: a hybrid Convolutional Neural Network Transformer (CNNT) and a new ResNet50 with a Convolutional Block Attention Module (ResNet50 CBAM). The novel ResNet50 CBAM architecture efficiently facilitates spatial and channel wise feature learning using attention mechanisms, allowing the model to focus on anomaly specific regions while eliminating irrelevant background noise. Comprehensive experiments were conducted on BabyEar4k, an expert annotated, publicly released dataset containing 3,852 high resolution ear images across five clinically relevant classes of anomalies. To substantiate the effectiveness of the proposed approach, a comparison was drawn with the current state-of-the-art scheme proposed by Ren et al., showing significant strides in classification performance. The ResNet50 CBAM model achieved a classification accuracy of 96.25%, an F1 score of 0.9576, and an AUC-ROC of 0.984, outperforming the CNNT model, which reached an accuracy of 93.97% and an F1 score of 0.9395, as well as the benchmark baseline F1 score of 0.832. These results highlight the robustness of the model and its generalization capability across classes of anomaly. The proposed framework addresses some crucial shortcomings of previous works, including reliance on binary classification, insufficient subtype granularity, and inadequate preprocessing. Employing BM3D-based denoising and SMOTE-based class balancing enhances the quality and learning efficiency. In all, the methodology presents a straightforward, accurate, and scalable approach to AI-based neonatal ear anomaly screening, offering substantial potential for wide integration into clinical decision-support systems to facilitate early intervention, decrease professional workload, and improve overall pediatric outcomes.