ENHANCED ATTENTION RESIDUAL PARALLEL U-NETS (EARPU) FOR LUNG TUMOR SEGMENTATION

Precise identification and classification of diseased tissue and its adjacent healthy structures are vital in the diagnosis of conditions like lung cancer. Achieving a more accurate diagnosis necessitates a substantial amount of data. Yet, physicians often encounter challenges in manually analyzing extensive and intricate CT scan images to extract essential information. While UNet-based architectures have demonstrated superior performance in image segmentation compared to other deep learning architectures, challenges arise in segmentation accuracy due to the low resolution of medical images and insufficient data. In this research, we propose a novel architectural design that addresses these issues by integrating four parallel UNETs through an attention residual network. To enhance performance, this architecture focuses on slicing the single image as four quadrant images and processing them individually rather than the entire image as a whole. This approach allows our model to capture intricate features of the images, as each image slice undergoes independent convolution and deconvolution through four parallel UNets. Ultimately, adhering to the attention residual network architecture, the UNet outputs are merged in a manner that amplifies the features of the image associated with the output through a skip connection. The suggested architecture demonstrated superior performance in terms of Dice score, achieving 91% on LIDI-IRDC, 89% on LUNA16, and 89% on Kaggle, compared to using a conventional U-Net or other U-Net variants.