A novel method for glioma segmentation and classification on pre-operative MRI scans using 3D U-Nets and transfer learning

This research paper proposes a novel approach that harnesses deep learning techniques to address two critical objectives in brain tumor analysis: segmentation and classification. We have used a 3D U-Net architecture to acquire spatial relationships and accurately delineate tumor regions from MRI images. These masks are superimposed onto the original images generating a lucid visualisation of the tumorous areas. For classification of tumors into HGG (high-grade gliomas) and LGG (low-grade gliomas) we created a customized CNN model and utilised transfer learning with VGG16, ResNet50, InceptionV3, MobileNetV2 and DenseNet121 as the pre-trained models. This approach has been tested using the BraTS 2019 dataset which bears a testimony to its cutting-edge performance in both segmentation and classification tasks. The denoising procedure yields an impressive average Peak Signal-to-Noise Ratio (PSNR) value of 97.82 dB, ensuring the production of denoised images of exceptional quality. The 3D U-Net employed for mask segmentation demonstrates precise delineation with a mean Intersection over Union (IoU) value of 0.97. Our custom CNN model exhibits exceptional efficacy, attaining training, validation and testing accuracies to be 97.99%, 97.99%, and 97.14% respectively. The proposed model has been compared with some of the state-of-the-art techniques and it has been found to outperform them as well. These findings underscore the model’s resilience, dependability, and potential to enhance brain tumor analysis, thereby facilitating accurate diagnosis and treatment planning.