Artificial Intelligence Reading of Postoperative Cochlear Implant X-Rays

Background/Objectives: Postoperative X-ray is the standard method for confirming cochlear implant (CI) electrode position, but interpretation varies between readers and subtle abnormalities can be missed. We evaluated a proof-of-concept deep learning framework for automated classification of CI electrode position on postoperative X-rays across four CNN architectures, each trained with and without class weighting, as an initial step toward determining whether such models warrant further clinical development. Methods: We analyzed 1033 radiographic regions of interest from 673 patients at a tertiary referral center. ResNet18, EfficientNet-B0, DenseNet121, and EfficientNet-B3 were evaluated using patient-level data splitting, transfer learning, and consensus ground truth from two independent specialists, assessed on an independent test set and through five-fold patient-level cross-validation, with Grad-CAM used to evaluate model attention. Results: DenseNet121 (non-weighted) achieved the highest AUC (0.932; 95% CI 0.836–1.000) and specificity of 98.6% (95% CI 95.3–99.6%), while EfficientNet-B3 (weighted) reached the highest sensitivity (69.2%; 95% CI 42.4–87.3%). Under cross-validation, ResNet18 (weighted) was the most consistent model (AUC 0.861 ± 0.030); single-split AUC values were optimistic by a mean of 0.069 across models. Grad-CAM confirmed electrode-focused attention in ResNet18 and DenseNet121. Conclusions: Multi-architecture CNNs showed potential to classify CI electrode position with high specificity and clinically interpretable attention, supporting further investigation as screening-support tools rather than autonomous diagnostic systems.