Artificial Intelligence-Assisted Inner Ear Computed Tomography Analysis: Radiomics-Based Comparison of Affected and Unaffected Ears in Idiopathic Sudden Sensorineural Hearing Loss.
Authors
Affiliations (5)
Affiliations (5)
- Department of Biomedical Engineering, College of Health & Sci., Gachon University, 1342, Seongnam-Daero, Sujeong-Gu, Gyeonggi-Do, 13120, Republic of Korea.
- Department of Otorhinolaryngology-Head and Neck Surgery, Gil Medical Center, Gachon University College of Medicine, 21 Namdong-Daero 774Beon-Gil, Namdong-Gu, Incheon, 21565, Republic of Korea. [email protected].
- Department of Biomedical Engineering, College of Health & Sci., Gachon University, 1342, Seongnam-Daero, Sujeong-Gu, Gyeonggi-Do, 13120, Republic of Korea. [email protected].
- Department of Biomedical Engineering, College of Medicine, Gil Medical Center, Gachon University, 38-13 3Beon-Gil, Namdong-Gu, Incheon, 21565, Republic of Korea. [email protected].
- Department of Health Science & Technology, Gachon Advanced Institute for Health Science and Technology, Lee Gil Ya Cancer and Diabetes Institute, Gachon University, 155 Gaetbeol-Ro, Yeonsu-Gu, Incheon, Republic of Korea. [email protected].
Abstract
We aimed to determine whether computed tomography (CT)-based radiomic features of the inner ear can distinguish the affected side from the contralateral normal-hearing side in patients with idiopathic sudden sensorineural hearing loss (ISSNHL) using a fully automated three-dimensional (3D) segmentation model. Deep learning-based inner ear segmentation followed by radiomics was hypothesized to reveal subtle structural differences associated with ISSNHL. This retrospective study included 318 patients who underwent 420 temporal bone CT scans. An independent test set consisted of 42 inner ear volumes from 21 patients with unilateral ISSNHL, including affected and contralateral normal-hearing sides. Inner ear structures were manually annotated by experienced otologists. A SwinUNETR-based 3D segmentation model was trained using an 8:1:1 dataset split with on-the-fly augmentation. Segmentation performance was evaluated using the Dice similarity coefficient(DSC), Intersection over Union (IoU), accuracy, precision, recall, and F1-score. A total of 1316 radiomic features were extracted from the automatically generated segmentation masks using PyRadiomics, encompassing original, wavelet, and Laplacian of Gaussian (LoG) derived feature classes. Group differences between the ISSNHL-affected and normal-hearing sides were assessed using Welch's t-test or the Mann-Whitney U test, with false discovery rate (FDR) correction applied. The segmentation model demonstrated high and stable performance, with comparable accuracy for ISSNHL-affected and normal-hearing sides. No significant differences were observed in radiomic features between the groups after correction. Principal component analysis and uniform manifold approximation and projection revealed no distinct clustering; the nine shape features with the lowest p-value features exhibited overlapping distributions. CT-based morphological radiomic features did not identify measurable structural differences between the affected and contralateral sides in ISSNHL, supporting the functional or microstructural nature of its underlying pathophysiology. Although automated 3D segmentation using SwinUNETR achieved highly accurate inner ear delineation, CT-derived radiomics demonstrated limited discriminatory value for ISSNHL. Alternative imaging biomarkers, functional imaging approaches, or deep learning-based representation features may be necessary for etiological assessment or prognostication.