Reswave-Net: A wavelet based Residual U-Net for Brain Tumour Segmentation and Overall Survival Prediction

A brain tumour is an abnormality in brain tissue that may cause harm to the nervous system and in severe cases can lead to death. Being a disease with a high mortality rate, the detection and accurate segmentation of brain tumour sub-regions is a crucial task in the disease diagnosis and treatment procedure. The manual segmentation process requires anatomical knowledge, is expensive, time-consuming, and inaccurate due to human errors. So, automatic and reliable segmentation methods are required; however, the large spatial and structural variability among brain tumours make automatic segmentation a challenging problem. This work proposes Reswave-Net, a deep learning network using an encoder-decoder (U-Net) architecture with residual connections to automate and standardize the task of tumour segmentation, which also incorporates wavelet decomposition of the input images. The network is trained and evaluated on Brain Tumour Segmentation (BraTS) Challenge-2020 dataset and achieves a mean Dice Score of 87.36%, 70.45%, and 72.55% and the Hausdorff distance of 6.87, 34.16 and 23.42 for the whole tumour, enhancing tumour and tumour core, respectively. For overall survival prediction, a random forest model is used where the radiomic features extracted from the image and age of the subject are used for training. The model achieves an accuracy of 58.4%.