Efficient Real-Time 3D Scene Reconstruction of Brain Tumors Using Convolutional Neural Networks and Image Processing Pipelines

3D scene reconstruction becomes more tedious to reconstruct the image accurately without losing pixels. This process mainly used to improve the accuracy of disease detection rate, surgical intervention, and neuro-navigation. In this paper, the proposed approach is an integrated model that combines various models such as novel Deformable Attention Transformers (DAT) and Hierarchical Latent Space (HTS), combined with hybrid image processing pipelines. In this work, preprocessing plays a significant role in improving algorithm performance. Various preprocessing techniques, such as Median Filtering, Cropping, ROBEX (Robust Brain Extraction), and Multi-Scale Edge Detection (Marr-Hildreth), are used to increase the data quality. In the next step, the Swin Transformer (Shifted Window Transformer) architecture is used to fine-tune the segmented tumor regions, which helps reconstruct 3D images. To ensure real-time performance, we integrate optimized data handling mechanisms and parallel processing strategies. Finally, the proposed approach improves the 3D reconstruction brain tumor images by using Pre-trained Fine-Tuned 3D U-Net model to obtained the missed patterns from the input images and transmit to the proposed DAT-HTS using transfer learning. The constructed 3D models are validated with ground-truth data by measuring Dice Similarity Coefficient (DSC), and Intersection over Union (IoU), and computational efficiency-based standard metrics. Experimental results show that the proposed method can provide high segmentation accuracy and fast reconstruction time, which can be used for clinical purpose. This will allow for more efficient surgical intervention planning, as well as better analysis of surgery results by visualization of deep brain tumor structures.