AI scoring tool shows promise in flagging cancerous skin lesions

Mayo Clinic researchers report that an artificial intelligence tool may give dermatologists additional information to help judge whether skin lesions are cancerous and to inform decisions about whether to biopsy or monitor them. The tool, called Dermoscopy Explainable Intelligence (DEXI), was evaluated in a retrospective study published in Mayo Clinic Proceedings, and its developers emphasize that it is intended to support, not replace, clinical judgment.

DEXI is part of an advanced imaging system that dermatologists use to capture and track skin lesions over time. The software assigns each lesion a score from 0.0 to 10.0 based on image characteristics associated with cancer risk. Beyond the number itself, DEXI uses explainable AI to identify which visual features, such as asymmetry, border and color, contributed to the score. This is meant to give dermatologists context they can weigh alongside their own clinical evaluation rather than presenting an opaque output.

The tool is a deep-learning system trained on more than 66,000 annotated dermoscopy images. In addition to producing a risk score, the explainability component indicates which image features drove the result, a design choice aimed at making the tool’s reasoning more transparent to clinicians.

The study included 178 patients who had undergone total-body photography, sequential digital dermoscopy imaging, and subsequent biopsy of one or more skin lesions. The researchers analyzed 261 lesions that had both a DEXI score and a diagnosis confirmed by tissue examination. They compared scores for noncancerous and cancerous lesions and further separated the cancerous lesions into melanoma and nonmelanoma skin cancers.

Using a threshold of 3.0, meaning a score of 3.0 or higher was considered positive for potential skin cancer, DEXI correctly identified 93.4% of cancerous lesions and 75% of noncancerous lesions. The average DEXI score was 2.5 for noncancerous lesions, compared with 7.8 for nonmelanoma skin cancers and 6.0 for melanomas. Cancerous lesions had significantly higher DEXI scores than noncancerous lesions overall. Based on these findings, the researchers identified 3.0 as a potential threshold that could help inform clinicians’ decisions about monitoring or biopsy.

“We identified 3.0 as a potential DEXI score threshold that could provide useful information when dermatologists are considering whether to monitor a lesion or perform a biopsy,” said Dr. Jenny Link, a dermatologist at Mayo Clinic and the study’s lead author. She noted that before the study she would review a patient’s DEXI score without knowing exactly what to do with the number, whereas she can now use the score as another piece of data to inform her judgment on whether to recommend a biopsy or monitoring.

The researchers also identified important limitations. DEXI does not account for patient-specific factors such as age, medications and history of melanoma, all of which can influence how a dermatologist weighs the risk of an individual lesion. The tool was also less effective at evaluating severely atypical nevi, or skin moles, and early evolving melanomas.

“Our study demonstrated that DEXI is a useful and valuable tool to support clinicians as they evaluate cancerous and noncancerous skin lesions, but it should not be used in isolation,” Dr. Link said. “The score must be interpreted alongside clinical findings and patient-specific factors, and the tool requires further validation and refinement.” Because the study was retrospective and limited to lesions with biopsy-confirmed diagnoses, the findings describe how the tool performed on this evaluated set rather than establishing how it would perform as a standalone screening method in routine care.

Subject of Research: Evaluation of an explainable AI dermoscopy tool for assessing skin cancer risk in skin lesions

Article Title: AI tool may support dermatologists in assessing skin lesions for cancer risk

Article References: AI tool may support dermatologists in assessing skin lesions for cancer risk. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, dermatology, dermoscopy, skin cancer, melanoma, nonmelanoma skin cancer, Mayo Clinic, deep learning, explainable AI, biopsy decision, retrospective study, Mayo Clinic Proceedings