Brain MRI 3D Segmentation using Patch Spatially Localized Network Tiles utilizing Pipelines on GPU Memory

Background and objective Quantitatively detailed WBTS is a very important role in brain image analysis and non-invasive automatic identify of brain regions. Massive input training models represent a significant challenge to the development of CNNs to purpose a large MRI image and analysis in computational neuroscience. DCNN has been widely applied to WBTS and GPU memory limitations for studying 2D/3D CNN methods are based on down-sampling. However, 3D patch-based high-resolution 3D-CNN has achieved superior performance and high accuracy by performing model training and big data learning. However, GPU memory is limited, and training time is longers. Methods In study, propose a customized, sub-optimal “patch-based are able to improve performance and optimize input channels. The GPU memory collected to carrying out the training process and training model inference and a collective larger than the activation size on pipeline. A sub-optimal “patch network tiles” modeling approach is needed for small input data. The PSLNT method distributes data into several independent 3D-FCN, so that WBTS obtains high-resolution, each network learns contextual information about fixed spatial locations. Results The proposed method through extensive experiments on 316 MRI brats 2021. calculating the average score, standard deviation, mean-dice score range wt 0.785, tc 0.761, and et 0.754 for intact tumors, tumor cores, and adjunctive tumors, also, as we can see the specificity metrics are wt 0.841, tc 0.825, and et 0.873 which means two segmentation images are the same for this metric. the metric average sensitivity was similar to approximately 98% for the entire tumor, tumor nucleus, and tumor enhancers respective. Conclusion Multi data patch network tiles with PSLNT produce improved segmentation results using 3D CNN. 3D U-net has the advantage of sens, spec, and HD, namely 0.78-0.98 and the difference between DSC and HD measurements is 0.11%. Results had slight advantages across the data set compared to all other methods evaluated in this study.