Brain Tumor Segmentation and Classification Using ResNet50 and U-Net with TCGA-LGG and TCIA MRI Scans

Abstract

Brain tumors have become a major source of death in the world. In the case of brain tumor, the brain cells of that particular part grow without any control. The growth has such a serious impact on the normal and healthy cells around the affected part of the brain. Malignancy and benignity are the two types of tumors. Symptoms of the tumor vary according to the place, size, and nature of the tumor. The variable nature of brain tumors is of such a complicated structure that it presents a big challenge for the academics in the field in terms of detection and early classification. A CNN-based model with enhanced “ResNet50 and U-Net architectures” was proposed in this paper. It was used in performing the required analyses on the publicly available “TCGA-LGG and TCIA datasets”. The data in the utilized datasets of “TCGA-LGG and TCIA included that of 120 patients”. The proposed CNN is used, combined with the fine-tuned ResNet50 model for detecting and classifying tumor versus non-tumor images. The model incorporates the U-Net model to precisely segment the tumor region. Accuracy, Intersection over Union (IOU), Dice Similarity Coefficient (DSC), and Similarity Index (SI) metrics are used for measuring the realization of the model. The quantitative results of fine-tuned “ResNet50 report IOU: 0.91, DSC: 0.95, and SI: 0.95”. The combination of U-Net with ResNet50 yielded the best of all, segmenting and classifying tumor regions effectively.