A lightweight asymmetric U-Net framework for acute ischemic stroke lesion segmentation in CT and CTP images

• Published in: Computer methods and programs in biomedicine 2022-11, Vol.226, p.107157-107157, Article 107157
• Main Authors: Kumar, Amish, Ghosal, Palash, Kundu, Soumya Snigdha, Mukherjee, Amritendu, Nandi, Debashis
• Format: Article
• Language: English
• Subjects: Computed Tomography; Computed tomography perfusion; Convolutional neural network; Humans; Image Processing, Computer-Assisted - methods; Ischemic Stroke; Neural Networks, Computer; Segmentation; Stroke - diagnostic imaging; Stroke lesion; Tomography, X-Ray Computed - methods
• ISSN: 0169-2607; 1872-7565
• DOI: 10.1016/j.cmpb.2022.107157

•A patch-based, residual, asymmetric, encoder-decoder 2D CNN architecture with 11 convolutional layers and 84,217 trainable parameters is introduced.•Utilized a training strategy combining the Focal Tversky and Binary cross-entropy loss functions to overcome the class imbalance issue in CT images.•A voting mechanism is applied to the result predicted by the weight matrices to ensure stable results.•Compared the method’s performance on the testing dataset with that of selected high-performing methods from the ISLES 2018 challenge and achieved the second rank. Background and objectives: This paper has introduced a patch-based, residual, asymmetric, encoder-decoder CNN that solves two major problems in acute ischemic stroke lesion segmentation from CT and CT perfusion data using deep neural networks. First, the class imbalance is encountered since the lesion core size covers less than 5% of the volume of the entire brain. Second, deeper neural networks face the drawback of vanishing gradients, and this degrades the learning ability of the network. Methods: The neural network architecture has been designed for better convergence and faster inference time without compromising performance to address these difficulties. It uses a training strategy combining Focal Tversky and Binary cross-entropy loss functions to overcome the class imbalance issue. The model comprises only four resolution steps with a total of 11 convolutional layers. A base filter of 8, used for the residual connection with two convolutional blocks at the encoder side, is doubled after each resolution step. Simultaneously, the decoder consists of residual blocks with one convolutional layer and a constant number of 8 filters in each resolution step. This proposition allows for a lighter build with fewer trainable parameters as well as aids in avoiding overfitting by allowing the decoder to decode only necessary information. Results: The presented method has been evaluated through submission on the publicly accessible platform of the Ischemic Stroke Lesion Segmentation (ISLES) 2018 medical image segmentation challenge achieving the second-highest testing dice similarity coefficient (DSC). The experimental results demonstrate that the proposed model achieves comparable performance to other submitted strategies in terms of DSC Precision, Recall, and Absolute Volume Difference (AVD). Conclusions: Through the proposed approach, the two major research gaps are coherently addressed while achieving high