Deep learning for volumetric assessment of traumatic cerebral hematoma

Abstract

Objective To develop a deep learning method for volumetric assessment of traumatic intracerebral hemorrhage (TICH) using the Trans-UNet model and to compare its performance with traditional formula-based methods. Methods CT data from 141 TICH patients admitted to Army Medical Center of PLA between May 2018 and May 2023 were collected. A deep learning method based on the Trans-UNet model was established. Manual delineation via picture archiving and communication system (PACS) was served as the gold standard for comparing the accuracy, consistency, and time efficiency of our method against 10 different formula-based methods for measuring the amount of TICH. Results The median volume of TICH, as manual delineation via PACS, was 1.167 mL, with a median measurement time of 135 s per patient. The median percentage error in volume between the deep learning method and manual delineation via PACS was 3.59%. Spearman correlation coefficient was 0.999 (P < 0.001), and a median measurement time was only 4.38 s per patient. In contrast, in the formula-based methods, the lowest median percentage error in volume was 16.451%, the highest Spearman correlation coefficient was 0.986 (P < 0.001), and the lowest median measurement time was 20 s for a single patient. The statistical differences were observed in percentage error in volume and measurement time between the 2 types of methods (all P < 0.001). Conclusion Our developed deep learning method for volumetric assessment of TICH is superior to the formula-based methods in terms of measurement accuracy and time efficiency.

Keywords

• traumatic intracerebral hemorrhage
• computed tomography
• deep learning
• volumetric quantification
• traumatic brain injury