(124) Deep Learning Method for Rapid Simultaneous Multi-structure Temporal Bone Segmentation

Introduction: The temporal bone (TB) contains multiple small but vital structures for which accurate segmentation can improve pre- and intra-operative guidance. This study presents a deep learning (DL) model for rapid automated segmentation of 9 key TB structures from CT scans.

Methods: 325 clinical TB CTs from a tertiary care hospital were used in this study. In each CT 9 key structures (Internal Carotid Artery [ICA], Chorda Tympani [ChT], Inner Ear [IE], External Auditory Canal [EAC], Facial Nerve [FN], Internal Auditory Canal [IAC], Ossicles [Oss], Otic Capsule [OtC] and Sigmoid Sinus [SS]) were manually segmented by an expert and used as a multilabel ground truth. A state-of-the-art DL algorithm (Swin UNETR) was adapted and trained to build a prediction model using 80% of the sample. The remaining 20% of the sample was used as a test set to evaluate the model using metrics such as Dice, Balanced Accuracy (BAcc), Volume similarity (VS), Average Symmetric Surface Distance (ASSD) and 95th Hausdorff Distance (HD95).

Results: On the test set, the metrics for all structures were [0.88; 0.95; 0.94; 0.18mm; 0.69 mm – Dice; BAcc; VS; ASSD; HD95 respectively]. Specifically, for the ICA [0.89; 0.98; 0.92; 0.30; 1.07] , for ChT [0.64; 0.83 ; 0.86; 0.21; 0.77], for IE [0.96; 0.99; 0.98; 0.05; 0.25], for EAC [0.93; 0.97; 0.97; 0.21; 0.92], for FN [0.83; 0.94; 0.91; 0.14; 0.56], for IAC [0.93; 0.98; 0.95; 0.14; 0.56], for Oss [0.90; 0.95; 0.96; 0.11; 0.44], for OtC [0.96; 0.98; 0.97; 0.06; 0.25], for SS [0.88; 0.94; 0.97; 0.38; 1.56]. Mean processing time was 9.1s per study.

Conclusions: A large training dataset and the use of optimized algorithms resulted in a robust model for fast simultaneous segmentation of key structures of the TB from clinical CT. This next generation of artificial intelligence application will provide valuable new datasets for otologic surgical planning and navigation.