MuRL: Multi-Resolution Deep Reinforcement Learning for Melanoma Diagnosis From Dermoscopic Images

Accurate and efficient melanoma diagnosis is essential for AI-driven clinical decision-making in dermatology. However, efficient classification of melanoma remains challenging due to the high computational resources required to achieve accurate diagnostic performance. In this paper, we propose MuRL, a reinforcement learning framework designed for accurate and efficient melanoma diagnosis from dermoscopic images. Specifically, MuRL integrates informative features extracted from multiple spatial resolutions to enhance feature representation by combining global context and fine-grained details for melanoma diagnosis. Furthermore, we propose a recurrent lesion-aware region localization that iteratively identifies salient regions within dermoscopic images, which enhances feature learning by focusing on diagnostically relevant areas. To reduce the overall computational cost for melanoma diagnosis, we propose a dynamic inference strategy that adaptively terminates the inference process. Experimental analyses on three publicly available datasets demonstrate the effectiveness and efficiency of our proposed method for melanoma diagnosis from dermoscopic images.