YARU3DFPN LIGHTWEIGHT MODEL TRAINING STRATEGY WITH EXTENSIVE TRAINING DATA FOR OBTAINING A BRAIN TUMOR SEGMENTATION GENERALIZED MODEL
Agus Subhan Akbar (a), Ahmad Hayam Brilian (b), Chastine Fatichah (c*), Alzena Dona Sabilla (d)

a,d) Universitas Islam Nahdlatul Ulama Jepara, Jepara, Jawa Tengah, Indonesia
b) Dinas Komunikasi dan Informatika Bojonegoro, Jawa Timur, Indonesia
c) Institut Teknologi Sepuluh Nopember, Surabaya, Jawa Timur, Indonesia

Abstract

The generalization capability of brain tumor segmentation models poses significant challenges in handling variations in the number of lesions per scan, lesion size and location within the brain, differences in MRI scanners and protocols used, and demographic factors such as age and gender. BraTS- GoAT provides an extensive training dataset to address these challenges. However, training models on large datasets re- quires substantial computational resources to achieve the de- sired segmentation and generalization performance. This re- search proposes using the Yaru3DFPN architecture with a data partitioning strategy to train multiple models. The train- ing dataset is divided into five parts, and each is used to train a separate Yaru3DFPN model. Each model is trained for 100 epochs, resulting in five segmentation models. Ensembling the five models to segment the validation dataset achieves lesion-wise dice scores of 68.95, 74.85, and 70.28 for ET, TC, and WT areas, respectively. These results demonstrate promising performance for further development.

Keywords: Yaru3DFPN, Lightweight Model Train- ing Strategy, Brain Tumor Segmentation, Ensemble Method

Topic: Deep Learning