RESUMO
BACKGROUND AND PURPOSE: To examine the diagnostic performance of unsupervised deep learning using a 3D variational autoencoder (VAE) for detecting and localizing inner ear abnormalities on CT images. METHOD: Temporal bone CT images of 6663 normal inner ears and 113 malformations were analyzed. For unsupervised learning, 113 images from both the malformation and normal cases were used as test data. Other normal images were used for training. A colored difference map representing differences between input and output images of 3D-VAE and the ratio of colored to total pixel numbers were calculated. Supervised learning was also investigated using a 3D deep residual network and all data were classified as normal or malformation using 10-fold cross-validation. RESULTS: For unsupervised learning, a significant difference in the colored pixel ratio was seen between normal (0.00021 ± 0.00022) and malformation (0.00148 ± 0.00087) cases with an area under the curve of 0.99 (specificity = 92.0%, sensitivity = 99.1%). Upon evaluation of the difference map, abnormal regions were partially and not highlighted in 7% and 0% of the malformations, respectively. For supervised learning, which achieved 99.8% specificity and 90.3% sensitivity, a part of and no abnormal regions were highlighted on interpretation maps in 34% and 8% of the malformations, respectively. Abnormal regions were not highlighted in 4 malformation cases diagnosed as malformations and were highlighted in 6 cases misdiagnosed as normal. CONCLUSIONS: Unsupervised deep learning of 3D-VAE precisely detected inner ear malformations and localized abnormal regions. Supervised learning did not identify whole abnormal regions frequently and basis for diagnosis was sometimes unclear.
