Deep learning for discriminating cochlear malformations on temporal bone CT

OBJECTIVES: Diagnosis of cochlear malformation on temporal bone CT images is often difficult because the imaging findings are frequently subtle. Our aim was to assess the utility of deep learning analysis in diagnosing cochlear malformation on temporal bone CT images. METHODS: A total of 373 temporal bone CTs (187 with normal and 186 with malformed cochlea) were included in this study. The network Swin UNETR was used for cochlea automatic segmentation, then we divided it into left and right sides. three classification networks, ResNet50, EfficientNet-B0 and DenseNet121, were used for training to diagnose cochlear malformation (590 sides). The test set (149 sides) was analyzed using deep learning models and by otologists. RESULTS: The average Dice coefficient of Swin UNETR was 0.93. The AUCs for ResNet50, EfficientNet-B0 and DenseNet121 were 0.93, 0.89 and 0.93, respectively, with DenseNet121 achieving the highest AUC among the four models. ResNet50 and DenseNet121 are better than otologists (1‒5 y). CONCLUSION: The deep learning technique may be a useful supportive tool for diagnosing cochlear malformation on temporal bone CT images.