ResCAM-Net: Residual Convolution and Atrous Multi-Scale Network for Robust Medical Image Segmentation

Abstract

Digital data and automated solutions have raised the standard of medical image analysis and made it more complex and challenging. Manual examination of medical images is inefficient, subjective, and error-prone due to variation in attributes such as shape, size, and texture. The majority of approaches in the literature, starting from the classical U-Net including its modern variants, usually suffer from overfitting issues, model saturation issues, and inability to capture multi-scale contextual features, reducing generalization across different imaging modalities. To address these challenges, we present ResCAM-Net, an improved encoder--decoder architecture that accurately and efficiently raises segmentation accuracy for different types of medical images. Our model incorporates residual learning, multi-kernel residual convolutions, adaptive feature recalibration using atrous spatial pyramid pooling, and a hybrid triple attention module to enhance feature aggregation and focus on critical regions. ResCAM-Net reduces the number of trainable parameters to 9.91 million, compared with other state-of-the-art architectures, by approximately 70 %, which significantly improves computational efficiency and reduces convergence time. Our model performs better on several benchmark datasets: Dice similarity coefficients of 89.84 %, 87.39 %, and 85 % on ISIC-2017 (small), ISIC-2017, and Kvasir-SEG datasets, respectively. Further, the robustness and generalization capabilities of ResCAM-Net were well reflected in these segmentation tasks, where it outperformed existing models both in accuracy and parameter efficiency.