Multi-class semantic segmentation and quantification of computed tomographic lesions in traumatic brain injury using deep learning–a multi-centre validation study.

Background: Computed tomography (CT) is the most common imaging modality in traumatic brain injury (TBI). However, its conventional use requires expert clinical interpretation and does not provide detailed quantitative outputs, which may have prognostic importance. Deep learning could reliably and efficiently detect, distinguish, and quantify different lesion types, providing opportunities for personalised treatment strategies and clinical research. Methods: An initial convolutional neural network (CNN) was trained and validated on expert manual segmentations (97 scans). This CNN was then used to automatically segment a new set of 839 scans, which were then manually corrected by experts. From these, a subset of 184 scans was used to train a final CNN for multi-class, voxel-wise segmentation of lesion types. The performance of this CNN was evaluated on a held-out test set with 655 scans. External validation was performed on a large, independent set of 500 patients from a different continent. Findings: When compared to manual reference, CNN-derived lesion volumes showed a mean error of 0·86mL (95% CI -5·23 to 6·94) for intraparenchymal haemorrhage (IPH), 1·83mL (-12·01 to 15·66) for extra-axial haemorrhage (EAH), 2·09mL (-9·38 to 13·56) for perilesional oedema and 0·07mL (-1·00 to 1·13) for intraventricular haemorrhage (IVH). Further, the CNN detected lesions with AUCs of 0·90 (0·86-0·94) for IPH, 0·80 (0·75-0·85) for EAH, 0·95 (0·89-1·00) for IVH on the external, independent patient dataset. Interpretation: We demonstrate the ability of a CNN to separately segment, detect and quantify multi-class haemorrhagic lesions and …