CHE-YU HSUChang, ChunhaoChunhaoChangWEI-WU CHENTsai, HsinhanHsinhanTsaiMa, ShihchiehShihchiehMaWEICHUNG WANG2023-05-172023-05-172022-01-01978303109001103029743https://scholars.lib.ntu.edu.tw/handle/123456789/631121The advancements in biotechnology and healthcare have led to the increasing use of artificial intelligence in medical imaging analysis. Recently, image recognition technology such as deep learning has become an important tool used for the detection and diagnosis of tumors. As labeling and annotation of tumors are time-consuming, it is necessary to design an approach that can automatically and accurately label tumors. Training a convolutional neural network (CNN) is possible to automatically interpret medical images more accurately, thereby assisting physicians in their diagnosis. In this paper, we describe an automated segmentation model by combining SegResnet and different loss functions to segment brain tumors in multimodal magnetic resonance imaging (MRI) scans and accelerate the tumor annotation process. By adding refinements to our training process, including region-based training, postprocessing, we were able to achieve Dice scores of 0.8159, 0.8734, and 0.9193, and Hausdorff Distance (95th percentile) of 20.02, 7.99, and 4.12 for the enhancing tumor (ET), whole tumor (WT), and tumor core (TC) respectively on the validation dataset.Boundary loss | Brain tumor | Dice loss | MRI scans segmentation | SegResnet[SDGs]SDG3conference paper10.1007/978-3-031-09002-8_302-s2.0-85135139431https://api.elsevier.com/content/abstract/scopus_id/85135139431