Identification of intracranial haemorrhage (ICH) using ResNet with data augmentation using CycleGAN and ICH segmentation using SegAN

Abstract

Intracranial Haemorrhage (ICH) occurring due to any injury to the brain is a fatal condition and its timely diagnosis is critically important. In this work, we propose a complete one-stop model for the identification of Intracranial Haemorrhage (ICH) and for the segmentation of ICH regions in Computerized Tomography (CT) images. The proposed method incorporates Residual Neural Network (ResNet) architecture for ICH identification and further segments the ICH region using an Adversarial Network called SegAN. This work incorporates a data augmentation method using CycleGAN, to solve the problem of class imbalance in the ICH dataset, leading to improved performance in the ICH identification task. CycleGAN is trained to convert a non-ICH CT slice into a synthetic CT slice with ICH, thereby augmenting the ICH sub-class, where there is a lack of data points. The proposed method achieved a macro average F1-score of 0.91 and a specificity of 0.99 and a sensitivity of 0.80 in the ICH identification task. Also, the proposed method works as a segmentation tool for all the five ICH sub-types and achieved a dice score of 0.32 and a mean Intersection Over Union (IOU) of 0.22. Thus, our proposed ICH identification and segmentation model can aid doctors in the accurate and timely diagnosis of ICH.