Deep Learning in Pediatric Ophthalmology: Differentiating Abusive Head Trauma in Retinal Fundus Images

Abstract

Purpose : To develop deep learning models for classifying abusive head trauma (AHT) in pediatric retinal fundus photographs.

Methods : This study analyzed a clinically validated dataset, including 192 AHT, 72 spontaneous vaginal delivery (SVD) with retinal hemorrhage (RH), and 1,000 normal post-delivery images captured via handheld devices from multiple institutions worldwide. Two deep learning models were utilized: a Vision Transformer (FastViT) and a Convolutional Neural Network (CNN), with focal loss applied to mitigate diagnostic imbalances. The dataset was randomly divided at the patient level in training (80%) and test (20%) sets. SHapley Additive exPlanations (SHAP) values were examined to improve the interpretability of the models' decision-making processes.

Results : The FastViT and CNN model achieved an overall accuracy of 98.9% and 97.7% on the test set, respectively. The FastViT model achieved a higher sensitivity (100%) than the CNN model (91.1%) due to the misclassification of 4 AHT test images. However, the CNN model had a higher specificity (100%) compared to the FastViT model (99.5%). SHAP analysis provided insights into feature importance, revealing key indicators used by the models in classification, thus enhancing understanding and trust in the models’ decisions.

Conclusions : The FastViT and CNN models demonstrate significant accuracy in differentiating between AHT and post-delivery (with and without RH) in fundus photos of pediatric patients. Utilizing SHAP values for model explainability further enriches their utility as supportive tools in clinical decision-making. While these models offer significant insights, their integration with the expertise of a multi-disciplinary medical team is essential to ensure a holistic, ethical approach in pediatric patient care.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

This table presents the comparative performance metrics of FastViT and CNN models in classifying pediatric retinal fundus photographs into three categories: AHT with RH, SVD with RH, and post-delivery (SVD or C-section) without RH.
Abbreviations: Abusive Head Trauma (AHT); Retinal Hemorrhage (RH); SVD (spontaneous vaginal delivery); Positive Predictive Value (PPV); Negative Predictive Value (NPV).

 

Confusion matrices for CNN (A) and FastViT (B) models across the three categories: Abusive Head Trauma (AHT) with retinal hemorrhage (RH), spontaneous vaginal delivery (SVD) with RH, post-delivery (SVD or C-section) with no RH.