Abstract / Summary
Medical imaging-based localization of malignant lesions is currently a challenging problem, in part because of the spatial coarseness of imaging systems and the sizeable computational cost of modern deep-learning systems, especially in resource-constrained edge-healthcare systems. Though the current developments in deep-learning have increased the accuracy of the diagnostics, the current state-of-the-art architecture is unable to maintain the fine-grained, high-resolution spatial detail and at the same time meet the real-time processing demands and low-power specifications posed by clinical practitioners. The current paper presents an overall overview of self-refined, attention-based lightweight deep-learning frameworks that were developed to overcome these shortcomings. We thoroughly study spatial (3D), channel, and hybrid attention mechanisms that are incorporated into highly efficient architectures, explaining their contribution to the enhancement of localization accuracy with a minimal computational cost. More importantly, the discourse involves a collection of methodological approaches, such as multi-scale feature learning, iterative attention refinement, and edge-based model optimization methods, such as pruning and quantization. Besides, the review also explores how different imaging modalities (including MRI, CT, histopathology, and thermal imaging) affect the effectiveness of attention-based lightweight models with a specific focus on their ability to reduce spatial coarseness in cancer localization in addition to their ability to detect tumors in real-time as part of next-generation edge-healthcare. Lastly, the paper discusses the existing challenges, which are scarce data, high annotation expenses, poor interpretability, and lack of clinical validation of the available datasets, hence the need to advance future studies to create strong, efficient, and clinically implementable cancer localization systems and to describe future research directions in this area.