A New Era in Oncologic Imaging Toward Early Detection of Melanoma: Artificial Intelligence and Beyond for Groundbreaking Multimodal Diagnostics

Abstract

Melanoma, the most lethal form of skin cancer, continues to represent a significant public health issue owing to its preferential metastasis and evolving incidence across populations. Even though there have been advances in the treatment of late-stage melanoma (PORTER et al., it is still associated with a poor prognosis [1-3], thus emphasizing the importance of early and accurate diagnosis. The diagnosis of melanoma is currently based on a diagnostic pathway featuring visual inspection, dermoscopy, histopathology confirmation, and molecular testing, but one that does so amidst inter-observer variability and delayed access to care because of resource-deprived environments, coupled with late stage at presentation. These limitations have encouraged the development of multimodal oncological imaging and AI-powered diagnostic solutions, as disruptive technological innovation in precision oncology. In this study, we provide general background on the integration of AI-based methods and multimodal imaging techniques for early melanoma diagnosis and characterisation, and discuss how these developments are expected to improve diagnostic accuracy and streamline clinical workflow. Within single modality-based techniques, dermoscopy, reflectance confocal microscopy (RCM), and optical coherence tomography (OCT) have shown a promising approach for non-invasive diagnosis of melanoma independently. However, each of these methods is also handicapped by trade-offs: ‘dermoscopy’ depends on operator skill, ‘RCM’ cannot penetrate very deep enough into the skin, and the cellular resolution of OCT lacks sensitivity for subtle differentiation. To circumvent these limitations, multimodal diagnostic integration became particularly appealing tools for the combination of morphological, chemical and functional information into a comprehensive tumour signature in recent studies. For instance it is able to visualize both the melanoma's structure itself and its selected S100B, HMB-45, and MITF biomarkers by means of fusing high-resolution optical imaging with molecular fluorescence mapping. Integrating these techniques into a single diagnostic system significantly increases the power for discrimination between benign nevi and early melanoma both in sensitivity and specificity. O the age of algorithmic models e.g. those involving machine learning (ML) and deep learning (DL), a new revolution in melanoma diagnosis has emerged. These models encompass complex deep-learning architectures (Convolutional Neural Networks such as ResNet, EfficientNet, and Vision Transformers), which are trained on large annotated image datasets to automatically detect suspicious lesions and quantify their likelihood of malignancy.