A Computer Vision Framework for Robust Medical Image Analysis Using Multi-Scale Deep Learning

Abstract

Accurate and reliable medical image analysis remains a challenging task due to significant variations in imaging modalities, acquisition conditions, and patient-specific anatomical differences. Conventional deep learning models struggle to generalize across such heterogeneous environments, particularly when confronted with noise, low contrast, small lesions, and ambiguous structural boundaries. To overcome these limitations, this study introduces RMS-DL, a Robust Multi-Scale Deep Learning framework designed to improve performance and consistency in real-world clinical applications. The framework incorporates a multi-scale depth-wise convolution module capable of capturing fine-grained textures and large contextual structures simultaneously, ensuring rich hierarchical feature extraction. Furthermore, a cross-scale dual attention mechanism combining channel attention and spatial attention—refines feature representations by highlighting clinically relevant regions while suppressing noise and irrelevant background artifacts. A lightweight multi-scale decoder aggregates multi-resolution features through attention-guided skip connections, enabling precise segmentation and stable classification outputs at low computational cost. Extensive experiments conducted across CT, MRI, X-ray, and ultrasound datasets demonstrate that RMS-DL consistently surpasses traditional CNN, attention-augmented architectures, and transformer-based models in Dice score, mIoU, boundary accuracy, and generalization capability. Additionally, the framework achieves significant reductions in FLOPs and parameter count, making it suitable for deployment in resource-constrained clinical environments. These results confirm RMS-DL as an efficient, robust, and scalable solution for next-generation medical imaging AI systems.