Brain Tumor Segmentation (BraTS) Challenge Short Paper: Improving Three-Dimensional Brain Tumor Segmentation Using SegResnet and Hybrid Boundary-Dice Loss
Journal
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Journal Volume
12963 LNCS
ISBN
9783031090011
Date Issued
2022-01-01
Author(s)
Abstract
The 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.
Subjects
Boundary loss | Brain tumor | Dice loss | MRI scans segmentation | SegResnet
SDGs
Type
conference paper