Efficient lung image segmentation with portable deep learning and panoptic attention mechanisms

Abstract

In biological image analysis, instance segmentation is crucial but challenging due to newlineoverlapping objects, blurry boundaries, complex backgrounds, and variations in object newlineappearances. This study introduces an enhanced DenseNet architecture I and#8722; DNet that newlineutilizes a residual attention feature integration method to effectively combine instance newlineand semantic information. I and#8722;DNet resolves segmentation challenges by augmenting newlinethe model s ability to assimilate contextual information through the integration of semantic newlinecontext and instance predictions. Reliability is enhanced by a robust sub-branch newlinethat aligns object recognition precision with predictive confidence. The research explores newlineportable deep learning methodologies for lung image segmentation, specifically newlinedesigned for X-ray imaging devices and independent of cloud computing. A novel newlinescaled attention mechanism s and#8722; AM is proposed, specifically designed for advanced newlineX-ray apparatus. A compilation of 360-degree lung X-ray images is utilized to train the newlines-AM model alongside other economical deep learning techniques. The rapid segmentation newlineapproaches are evaluated and contrasted using several accuracy metrics. A novel newlinepanoptic-segmentation model is introduced to markedly enhance segmentation accuracy newlineby bidirectional communication within a transformer network, integrating instance newlineand semantic information. The bidirectional link established between the memory path newlineand the CNN pixel path ensures the seamless operation of the dual-path transformer newlineblock. The system integrates multiscale features from various decoding algorithms using newlinestacked decoding blocks. The transformer architecture integrates essential mechanisms newlinesuch as pixel-to-pixel self-attention, memory-to-memory self-attention, pixelto- newlinememory feedback attention, and memory-to-pixel attention. The proposed models newlinedemonstrate significant improvements in lung segmentation tasks, enhancing both performance newlineand reliability. These advancements streamline the medical workflow and newlineimprove patient care outcomes by