Automatic melanoma detection and segmentation in dermoscopy images using deep RetinaNet and conditional random fields

Melanoma is one of the major causes of death around the world and is also known as malignant skin cancer. Melanoma detection is possible at an early stage by visual inspection of the infected lesions. There are a limited number of expert dermatologists available, moreover visual inspection also has limited accuracy. Hence, diagnosis and clinical decision making can be complicated for melanoma detection. Towards this, we propose a deep learning method for automatic detection and segmentation of melanoma regions within the dermoscopic images for precise melanoma segmentation. Our method generates bounding boxes around multiple regions to precisely detect the affected regions using RetinaNet. Further, conditional random field (CRF) is applied to the detected regions for segmentation of the melanoma lesion. In particular, we perform three steps: image pre-processing, melanoma localization, and melanoma segmentation. We evaluate our proposed method on Pedro Hispano (PH)2, International Skin Imaging Collaboration (ISIC) 2017, and ISIC 2018 benchmark datasets. Our experimental findings reveal the performance supremacy of our proposed method. For instance, pixel-level sensitivity is 0.932, pixel-level specificity is 0.977, pixel-level accuracy is 0.942, dice coefficient is 0.931, and Jaccard index is 0.9187 for ISIC 2018 challenge data. Our proposed method has shown good performance against other state-of-the-art methods evaluated on the ISIC 2018 challenge dataset. We attribute this performance to deep features computation using RetinaNet for detecting melanoma region and CRF for precise segmentation of the melanoma lesion.

Automated summary

Melanoma, a major cause of death worldwide, is challenging to diagnose and treat. This study proposes a deep learning method that combines image processing, localization, and segmentation to accurately detect and segment melanoma regions.