Fully automatic brain tumor segmentation using DeepLabv3+ with variable loss functions
Sakshi Ahuja, Bijaya Ketan Panigrahi, Tapan Kumar Gandhi
Abstract
Glioma is found to be the deadliest and rapidly growing brain tumor in adults. The proposed pixel-level-based segmentation framework along with combined loss resolves the class imbalance issue. The proposed methodology is implemented on FLAIR MRI modalities from the training dataset from Brain Tumor Segmentation 2020 (BraTS 2020) challenge. Initially, the input dataset is pre-processed using numerous operations such as resizing, and normalization, etc. The pre-processed dataset is split into 90% training set, (80% training set and 10% validation), and 10% of the testing set. The training set is augmented to reduce the data overfitting issue. The proposed methodology is implemented with DeepLab v3+ with different baseline networks such as ResNet18, and MobileNetv2. Also, the pixel classification layer is replaced with different loss functions to efficiently segment the small size tumors. Further, the proposed methodology with combined loss obtained an average dice index of 0.92±0.5, mean IOU of 0.93, global accuracy of 99.67%, and F1-score of 0.94±0.2.