Automated Chronic Subdural Hematoma Segmentation with 3D U-net

Introduction: Chronic subdural hematoma (cSDH), commonly caused by traumatic brain injury (TBI), is significantly more prevalent in veterans, with an incidence of 121.4 per 100,000 persons per year, compared to 1.72–20.62 in civilians. Management of cSDH heavily relies on computed tomography (CT) imaging, with serial scans guiding treatment. This retrospective study aims to develop a deep-learning-based machine learning (ML) model to automate cSDH segmentation, enhancing clinical diagnosis. The model is trained on a veteran cohort from the Veteran Affairs New York Harbor Healthcare System (NYHHS), considering population-specific factors in cSDH incidence analysis.

Methods: We obtained 65 CT scans from NYHHS Veteran Affairs Hospital, identified by CPT codes 6110-61108 for Subdural Evacuating Port System (SEPS) drainage. Hematoma volume ranged from 43.25ml to 484ml, with 10 cases of bilateral and the rest unilateral cSDH. Manual segmentation of cSDH, confirmed by neurosurgeons, was used as the ground truth. A 3D U-Net architecture was employed to automate the segmentation, utilizing these manually segmented CT scans. The model, trained using TensorFlow, includes contracting and expanding paths with convolutional and upsampling layers to capture multi-scale features for accurate cSDH segmentation.

Results: Our best performing model demonstrated remarkable efficacy, achieving a mean Intersection over Union (IOU) of 0.8743 and a Dice similarity coefficient (DICE) score of 0.9214 on the validation set. Cases where the model performed well typically exhibited clear and distinct segmentation of the cSDH region, closely aligning with the manually segmented ground truth. However, certain cases posed challenges for the model, particularly those with faintly defined hematoma boundaries. These instances often resulted in suboptimal segmentation, highlighting areas for potential model refinement and improvement.

Conclusion : The 3D U-Net model demonstrated high accuracy in delineating cSDH volumes, offering clinicians valuable diagnostic assistance and treatment planning support. Its successful implementation highlights its potential as a reliable, efficient tool for automated cSDH segmentation, promising improved patient care and clinical outcomes in managing this common neurological condition.