Optimizing Brain Tumor Segmentation: A Lightweight Yaru3dfpn Training Strategy with Extensive Data Partitioning

The generalization capability of brain tumor segmentation models presents significant challenges due to variations in lesion count, size, and location, as well as differences in MRI scanners, imaging protocols, and patient demographics such as age and gender. A key difficulty is balancing segmentation accuracy with computational efficiency, particularly when training models on large-scale datasets like BraTS-GoAT, which, despite offering extensive training data, require substantial computational resources. This study addresses these challenges by introducing a lightweight yet effective segmentation approach utilizing the Yaru3DFPN architecture combined with a data partitioning strategy. The training dataset was divided into five subsets, each used to train an independent Yaru3DFPN model for 100 epochs, resulting in five distinct segmentation models. By ensembling these models, segmentation on the validation dataset achieved lesion-wise Dice scores of 68.95, 74.85, and 70.28 for the enhancing tumor (ET), tumor core (TC), and whole tumor (WT) regions, respectively. These findings demonstrate the effectiveness of the proposed approach in improving segmentation performance while maintaining computational efficiency, making it a promising direction for further advancements in automated brain tumor diagnosis.