The AGU-Net Architecture for Brain Tumor Segmentation: BraTS Challenges 2023

For patients suffering from brain tumors, prognosis estimation and treatment decisions are made by a multidisciplinary team of medical doctors based on a set of MR scans. Currently, the lack of automatic, standardized, and robust methods for tumor characterization represents a major hurdle for use in clinical practice. This paper describes our contribution to the BraTS 2023 Continuous Evaluation challenge for the segmentation of all tumor types, using our single-stage AGU-Net architecture and various training strategies. Performance over the training sets were reported using our custom pixel-wise, patient-wise, and lesion-wise metrics. For the tumor core, an average lesion-wise Dice score of 85% was obtained over the glioma challenges, 80% for the meningioma challenge, and 73% for the metastasis challenge. Performance reported over the validation and test sets were officially computed by the challenge team. Over the test sets, an average lesion-wise Dice score of 75% to 80% was achieved for the tumor core over the glioma and meningioma challenges, while a lower score of 43% was reached for the metastasis challenge. The proposed method performed well and showed an ability to generalize on challenges with sufficient data.