Progressively Trained Deeply Supervised Network for Brain Tumor Segmentation

Glioblastoma, a highly aggressive brain cancer, poses significant diagnostic and treatment challenges. Magnetic Resonance Imaging (MRI) is the standard method for tumor detection, traditionally requiring manual annotation by radiologists. Recent advancements in deep learning have enabled automated segmentation and classification of brain tumors, improving efficiency and accuracy. However, deep learning models often demand substantial computational resources, limiting accessibility and slowing iteration cycles. This study proposes a progressive training strategy that decomposes large neural networks into sequentially trained segments, reducing GPU memory usage by up to 40% while maintaining competitive segmentation performance. Using a ConvNeXt encoder and a UNet3+ decoder, the approach was tested on the BraTS 2021 dataset. The results from our experiments demonstrated that the progressive training method achieved comparable Dice scores to end-to-end training for 3 tumor subregions: 82.39 for enhancing tumor (ET), 87.57 for tumor core (TC), and 90.80 for whole tumor (WT). The staged approach provides added flexibility for researchers to optimize hyperparameters and address implementation challenges early in the training process. While the method increases total training time, it facilitates efficient use of computational resources and offers practical advantages for resource-constrained environments.