Attention-based deep learning segmentation: Application to brain tumor delineation
Reza Karimzadeh, Emad Fatemizadeh, Hossein Arabi
Abstract
Brain tumor segmentation is an important step in brain cancer diagnosis and treatment. Manual segmentation is highly time consuming and tedious. To address these issues deep learning-base methods have been employed and demonstrated tremendous improvement in terms of performance accuracy and time-efficiency compared to the conventional segmentation methods. Due to varying size, location and shape of tumors, tumor segmentation task is highly challenging in comparison to the anatomical organ segmentation, thus it seems essential to develop robust, accurate, and efficient deep learning-based solutions for this task. In this work, we proposed a method that utilized UN et architecture as backbone (referred to as Attention-based UNet (AbUNet)) to establish an efficient tumor segmentation framework. AbUNet consist of two basic components, attention and segmentation modules. The purpose of attention module is to estimate approximate location of tumor through a rough/coarse segmentation with high true positive rate to guarantee the inclusion of the entire tumor. To this end, a new cost function was introduced and the dilated masks of ground-truths were employed for training of the attention module. The outcome and some feature maps from decoder part of attention module were fed to the segmentation module to guide the segmentation network focusing on the confined regions of the image which contain tumor. In the last step, the segmentation module predicts final mask of the tumor. To evaluate our proposed method, we used the BraTS dataset which contains brain MRI scans with manually defined tumor masks delineated by an expert. Our baseline UNet without attention module achieved a Dice score of 0.68 compared to AbUNet with Dice score of 0.79 that indicates significant improvement in segmentation accuracy. It was demonstrated that incorporation of the proposed attention module in the deep learning-based segmentation network would significantly enhance the robustness and accuracy of the segmentation.