INTEGRATION OF ILLUMINATION ENHANCEMENT APPROACHES FOR IMPROVING SKIN LESION SEGMENTATION AND MELANOMA DETECTION IN MACROSCOPIC IMAGES

Abstract

Skin cancer represents a significant global health concern, with melanoma being the deadliest type and accounting for 132,000 cases annually. Early detection is paramount, and while dermoscopy aids in revealing subsurface skin lesion structures, macroscopic imaging is increasingly vital due to its accessibility and ease of use. However, research on macroscopic images, remains limited due to challenges such as uneven background lighting and exposure, which worsen the performance of deep learning (DL) models in lesion segmentation and classification.

Addressing these challenges, this study introduces innovative supervised and unsupervised approaches to low-light image enhancement on macroscopic images. Specifically, we evaluate the performance of an unsupervised DL model, zero-reference deep curve estimation (Zero-DCE), which we have optimized using new loss functions, along with transfer learning and fine-tuning techniques. This model is then compared with state-of-the-art supervised methods. We aim to improve segmentation and classification processes on low-light macroscopic images through the application of the optimized Zero-DCE method and provide interpretability of the developed models.

In our approach, various segmentation architectures including U-Net, Attention U-Net, Dense U-Net, Dense Attention U-Net, and U-Net 3+ were trained and assessed using several metrics: Dice similarity coefficient (DSC), intersection over union (IoU), and the area under the curve - region of convergence (AUC-ROC) among others. For classification, we utilized models like EfficientNet, ResNet, MobileNet, and VGGNet in a binary classification framework to distinguish between melanoma and not melanoma lesions. Three public macroscopic image datasets¿Waterloo, PAD-UFES-20, and 7-point criteria¿were incorporated in our evaluation.

Our findings reveal that the best segmentation outcomes were achieved using images enhanced by optimized Zero-DCE in combination with Dense Attention U-Net, exhibiting DSC, IoU, and AUC-ROC of 88.8%, 79.9%, and 94.6%, respectively. In terms of classification of segmented images, we also achieved promising results: an AUC-ROC of 76.5%, 85.4%, and 68.7% on the Waterloo, PAD-UFES-20, and 7-point criteria datasets, respectively. Moreover, the application of interpretability methods, such as Gradient-weighted Class Activation Mapping, demonstrated that segmentation techniques are vital for focusing on relevant lesion characteristics during classification. Our research contributes to the development of more effective CAD systems for clinical decision-making in dermatology, with approaches capable of enhancing low-light on macroscopic images without the need for reference images.