Challenges and Future Prospects

The development and implementation of AI in skin cancer screening face several key challenges. One significant challenge is generalizability. Many image databases concentrate on a limited range of diagnoses and skin types, which can reduce an AI model's effectiveness in real-world clinical settings. These databases often lack comprehensive representation of various lesions, including less common or benign types. Additionally, there may be a bias toward patients who are more tech-savvy or have better access to healthcare facilities equipped with AI technology..

Another significant challenge is the scarcity of real-world validation studies to confirm the practical utility of AI and machine learning. Most research has been conducted in controlled experimental environments, underscoring the need for prospective studies in actual clinical settings to ensure that AI models are applicable and effective in practice.

Privacy concerns also arise from the handling and storage of sensitive patient data, including images and medical history. Most open-access databases are limited in the number of images they provide due to copyright and privacy issues. These databases often focus primarily on dermoscopic images and may lack clinical metadata and full-body photos, which could enhance diagnostic accuracy by providing additional context. Variability in imaging conditions, such as lighting and camera settings, along with the absence of standardized protocols, adds complexity to the situation. Additionally, the opacity of machine-learning algorithms can make it difficult for clinicians to understand the reasoning behind a diagnosis, potentially impacting trust and accountability.

International Skin Imaging Collaboration (ISIC)

Effective AI development requires extensive datasets. The ISIC Archive was created to address this. Sponsored by the International Society for Digital Imaging of the Skin (ISDIS), the International Skin Imaging Collaboration (ISIC) is a joint effort between academia and industry aimed at enhancing melanoma diagnoses and reducing mortality through digital imaging technologies. ISIC provides a public database for benchmarking machine learning algorithms and hosts public challenges.

ISIC activities are divided into two main areas:

ISIC Working Groups: These expert groups engage both the dermatology and computer vision communities to develop standards for skin imaging, including imaging technologies, techniques, terminology, and metadata, to improve the quality, privacy, and interoperability of digital skin imaging.
ISIC Archive: ISIC maintains the largest publicly available database of skin lesion images, collected from leading centers around the world, with over 40,000 images (primarily dermoscopic, but also clinical) of around 30 different skin lesions, each tagged with a diagnosis.

Since 2016, ISIC has hosted five Grand Challenges, all focused on developing AI for diagnostic classification using dermoscopy images. Participants are invited to submit their algorithms and compete for the most accurate algorithm. The challenges consist of three steps: (1) segmentation of the lesion from the background of the images; (2) detection of different dermoscopic features; and (3) classification of the lesion in the image. Each year, the challenges have become more complicated, with more images and additional diagnoses included. ISIC challenges have significantly increased the amount of research and publications related to AI and skin lesion diagnosis.

Algorithms that can accurately distinguish between benign and malignant skin lesions have the potential to improve skin cancer triage, especially in populations with limited access to specialized dermatologic care. Training algorithms on lower-quality images could enhance clinical workflows and enable earlier detection in primary care or non-clinical settings, where photos are often taken by non-expert physicians or patients. However, most public skin cancer datasets rely on dermoscopic images, which are limited by selection bias and lack of standardization, making them more suitable for skilled clinicians. The ISIC 2024 Challenge addressed these concerns and aimed to develop AI algorithms to differentiate between malignant and benign lesions. Hosted on Kaggle, the challenge featured a dataset of 400,000 skin lesion images from thousands of patients across three continents, including image crops resembling cell phone photos, using data from Three-Dimensional Total Body Photography (3D TBP).