Renal Medical Image Segmentation using Cross-Perceptron Deliberation and Multi- Scale Feature Fusion

Architectures based on Convolutional Neural Networks (CNNs), like U-Net, have demonstrated notable efficiency in the segmentation of renal medical images. However, because convolution processes are limited and have restricted accessible fields, they frequently have trouble capturing long-range dependencies. Recent developments have improved global context modeling by incorporating transformer modules into U-Net variations to address this issue. However, during the global fusion process, these transformers based methods run the risk of losing important local spatial information. This research introduces Multi-Scale MCPA, a unique architecture designed specifically for the segmentation of 2D renal medical images. An encoder, decoder and cross perceptron module are the three main parts of MCPATo provide rich multi-scale feature interaction, the Cross Perceptron primarily uses several Multi-Scale Cross Perceptron (MCP) modules to capture local dependencies. To efficiently model long-range dependencies, these features are spatially unfolded, concatenated, and processed by a Global Perceptron component. A Progressive Dual-Branch Structure (PDBS) is implemented to enhance segmentation performance, particularly for fine-grained structures. During training, this component guides the network to progressively transfer its attention from coarse structural elements to intricate pixel-level representations. The proposed method is specifically designed for 2D medical image segmentation tasks, given the clinical significance of 2D imaging and the high computing demands of 3D models. Experimentation of the proposed approach on multiple publicly accessible datasets from different imaging tasks and modalities, such as OCTA (ROSE), fundus images (DRIVE, CHASE_DB1, HRF), MRI (ACDC), and CT (Synapse), demonstrates that the proposed method reliably outperforms state-of-the-art segmentation methods, accomplishing enhancements of +2.1% Dice score on Synapse CT, +2.6% on ACDC MRI, and up to +3.4% on retinal fundus datasets. The effectiveness and generalizability of MCPA are established by experimental results, which show that it routinely outperforms existing techniques in segmentation accuracy.