Automatic Alignment and Performance Validation of Temporal Bone Standard Planes in U-HRCT

Objective: The present study aimed to validate an automatic standard plane alignment algorithm for temporal bone ultra-high resolution computed tomography (U-HRCT) using large-scale imaging datasets. Methods: Using lateral semicircular canal segmentation and statistical priors of angles between the reference lines and planes, the proposed algorithm solves the problem of tedious and inefficient manual post-processing during unilateral temporal bone examinations. We accordingly performed a clinical efficacy evaluation and analyzed cases of alignment failure to assess the reliability of this algorithm. A total of 3094 patients (5668 ears) with U-HRCT data were retrospectively enrolled. An inner ear substructure segmentation network model was used to automatically segment the lateral semicircular canal (LSC). Statistical priors of the extracted reference line-plane angles θ were adopted to identify and align the standard unilateral planes. A three-point scoring system was adopted to evaluate the automatic alignment outcomes of all 5668 ear samples. For samples with failed alignment, further three-point scoring was conducted on the corresponding automatic LSC segmentation results, followed by an analysis of the causes of failure. Finally, the LSC automatic segmentation scores of the randomly sampled success and failure groups were compared. Results: Among the 5668 enrolled ears, scoring results demonstrated that 69.83% (3958 ears) achieved a score of 3 (excellent), 21.81% (1236 ears) scored 2 (good), and 8.36% (474 ears) scored 1 (failure), yielding a qualified rate (score ≥2) of 91.64%. Among the 474 ears with alignment failure, none of the automatic LSC segmentations attained a score of 3 (intact segmentation with well-defined borders); 6.12% (29 ears) were graded as 2 (essentially intact segmentation with partial marginal deviation), and the remaining 93.88% (445 ears) were graded as 1 (incomplete segmentation, blurred boundaries, or positional offset). There was a statistically significant difference in LSC segmentation scores between the qualified and failed alignment groups, indicating that poor LSC segmentation is the predominant direct contributor to alignment failure. Etiological analysis of failures revealed the following constituent ratios: inner ear anatomical variation or malformation sparing the lateral semicircular canal (10.34%), abnormal temporal bone mineral density or osseous defects unrelated to LSC (27.00%), postoperative changes secondary to metallic implant placement (12.66%), and intrinsic algorithmic failure of automatic segmentation (50.00%). Conclusion: Overall, the proposed automatic alignment method, relying on LSC segmentation and statistical priors of reference line-plane angles, achieved a qualified rate of 91.64% for standard slice alignment of the unilateral temporal bone on U-HRCT, which markedly improved the standardization and efficiency of radiological post-processing. Alignment failure is strongly correlated with LSC segmentation quality, providing a clear direction for the optimization of subsequent algorithms.