TBI lesion segmentation in head CT: Impact of preprocessing and data augmentation

Miguel Monteiro, Konstantinos Kamnitsas, Enzo Ferrante, Francois Mathieu, Steven McDonagh, Sam Cook, Susan Stevenson, Tilak Das, Aneesh Khetani, Tom Newman, Fred Zeiler, Richard Digby, Jonathan P Coles, Daniel Rueckert, David K Menon, Virginia FJ Newcombe, Ben Glocker

Published in Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries. 5th International Workshop, BrainLes 2019, 2020

[paper] [cite] [poster]

Abstract

Automatic segmentation of lesions in head CT provides key information for patient management, prognosis and disease monitoring. Despite its clinical importance, method development has mostly focused on multi-parametric MRI. Analysis of the brain in CT is challenging due to limited soft tissue contrast and its mono-modal nature. We study the under-explored problem of fine-grained CT segmentation of multiple lesion types (core, blood, oedema) in traumatic brain injury (TBI). We observe that preprocessing and data augmentation choices greatly impact the segmentation accuracy of a neural network, yet these factors are rarely thoroughly assessed in prior work. We design an empirical study that extensively evaluates the impact of different data preprocessing and augmentation methods. We show that these choices can have an impact of up to 18% DSC. We conclude that resampling to isotropic resolution yields improved performance, skull-stripping can be replaced by using the right intensity window, and affine-to-atlas registration is not necessary if we use sufficient spatial augmentation. Since both skull-stripping and affine-to-atlas registration are susceptible to failure, we recommend their alternatives to be used in practice. We believe this is the first work to report results for fine-grained multi-class segmentation of TBI in CT. Our findings may inform further research in this under-explored yet clinically important task of automatic head CT lesion segmentation.