Oleh Dzyubachyk

LUMC

In short, Oleh Dzyubachyk's research entails the following:

Identifying Individuals at High Risk of Melanoma Using Total-Body Photographs: A Deep-Learning Approach

Melanoma is an aggressive type of skin cancer associated with significant mortality and healthcare costs. The lifetime risk of developing melanoma in the Netherlands is 1%, with certain subgroups being at higher risk due to hereditary susceptibility, multiple atypical nevi (moles), fair skin type and excessive exposure to ultraviolet radiation. Early diagnosis is critical to prevent metastatic dissemination, and periodic skin examinations are recommended for patients at high risk. Current criteria to identify patients who are advised to undergo lifelong periodic screening are crude and need refinement. The screenings are associated with considerable healthcare costs and are not feasible for all individuals at risk (estimated 1 million persons in the Netherlands).

We have collected a unique set of total-body photographs and detailed clinical data from patients at increased risk of melanoma. By leveraging the potential of deep-learning, the algorithm will identify those patients with atypical nevi and inherited melanoma susceptibility who are at a particularly high risk of developing melanoma. This will enable clinicians to select patients at high risk of developing melanoma for periodic skin examination, instruction for skin self-examination, sequential photography, and suspicious skin lesion removal, while reducing unnecessary skin cancer screening. Capacity of deep-learning algorithms to detect patterns associated with high melanoma risk in these multi-modal data, can, in turn, facilitate development of precision prevention strategies for those at high and at moderate risk of melanoma.

Short summary of progress/results June 2026

In this initial phase of the project, we identified important directions and milestones that would help us on the path towards reaching our ultimate goal in this project, namely to develop a multi-modal melanoma risk estimation system based on deep learning and total-body photography (TBP). In particular, we analyzed the availability of the clinical data and the corresponding annotations and discovered that dermoscopy images are much more abundant compared to the TBP. We consequently demonstrated that performing classification of skin lesions on the TBP data highly benefits from pre-training the model on the dermoscopy data. The corresponding manuscript has been (provisionally) accepted for presentation at MICCAI (Medical Image Computing and Computer Assisted Intervention; top-ranked conference in the medical image processing field).

Next, as a first step towards quantifying the total-body data, we established collaboration with expert clinicians and statisticians. We believe that proper problem formulation will be crucial for our project, as most of the subsequent tasks, including but not limited to data annotation and methodology development, will highly depend on the chosen problem formulation. In particular, to get a better perspective on this problem, we established collaboration with one of the leading experts in this field. A joint manuscript highlighting challenges of performing melanoma risk prediction with the help of AI models is currently in preparation.