Brain tumor segmentation using U-Net++ with dense connection

One of the most crucial tasks in the diagnosis and treatment planning of patients with brain cancer is the segmentation of brain tumor regions from MRIs. Conventional manual segmentation techniques take a lot of time and are prone to mistakes by people, leading to inconsistencies and inefficiencies in clinical practice. A convolutional neural network model identifies and classifies tumor regions within the brain using MRI scans. A more sophisticated variant of the U-Net design known as U-Net++ has layered and densely linked skip routes that enhance the network's capacity to record contextual and fine-grained information. Dense connections, which mitigate the vanishing gradient problem and encourage feature reuse, are incorporated into U-Net++ to further improve gradient flow and feature propagation. This architecture is made up of a sequence of blocks that are both encoders and decoders. The encoder route blocks extract progressively abstract information, while the matching decoder path blocks rebuild the input image's spatial resolution.The study utilizes the MICCAI BraTS2020 dataset for training and evaluation, which contains a collection of labeled brain tumor MRI scans. Various data augmentation techniques are employed to enhance the model's performance. The models' average training accuracy for whole tumor (WT) values is as follows: Simple U-Net, U-Net with Dropout, U-Net with Batch Normalization, U-Net with Res Path, and U-Net++ with Dense Connection are 66%, 50%, 99%, 100%, and 100%, respectively. Similarly, the average validation accuracy for whole tumor (WT) values is 65%, 50%, 98%, 99%, and 99%, respectively. Furthermore, the average DICE scores for whole tumor (WT) values for the models are 0.60, 0.70, 0.84, 0.90, and 0.93, respectively. Finally, the average Intersection Over Union (IOU) for whole tumor (WT) values for the models are 0.43, 0.55, 0.72, 0.82, and 0.86, respectively, on the brain tumor segmentation dataset BraTS2020.