AI Model Matches Dermatologist Expertise
A simple AI model has been shown to perform on par with experienced dermatologists when assessing the aggressiveness of squamous cell carcinoma, a common form of skin cancer. This groundbreaking research, led by the University of Gothenburg, sheds light on the increasing prevalence of this disease in Sweden, where more than 10,000 cases are diagnosed annually.
Squamous cell carcinoma is the second most common skin cancer in Sweden, following basal cell carcinoma. The disease predominantly arises in areas of the skin that have been extensively exposed to sunlight over the years. Associate Professor Sam Polesie, a dermatologist who spearheaded the study, explained that this cancer type results from mutations in the skin’s most common cells, which are heavily influenced by accumulated UV radiation. Lesions may present as rough, scaly patches with uneven pigmentation and reduced elasticity.
Although diagnosing squamous cell carcinoma is typically straightforward, the real challenge lies in preoperative assessment. Physicians must gauge how aggressively the tumor is growing to determine the urgency and extent of the necessary surgical intervention. More aggressive tumors necessitate prompt action and often require extensive tissue removal, while less severe cases can be managed with simpler procedures.
In Sweden and many other countries, preoperative punch biopsies are not routinely performed for suspected skin cancers. Instead, surgeries are generally conducted based on clinical impressions, with entire excised specimens analyzed histopathologically afterward. This practice highlights the need for alternative assessment methods, such as image analysis using artificial intelligence.
During the study, researchers trained an AI system utilizing 1,829 close-up images of confirmed squamous cell carcinoma cases. This model’s capacity to differentiate three levels of tumor aggressiveness was evaluated against assessments from seven independent dermatologists, yielding remarkably similar results. Notably, the study revealed that individual dermatologist assessments showed only moderate agreement, further emphasizing the complexities involved in this diagnostic challenge.
The research identified specific clinical features that may indicate more aggressive tumor growth, such as ulcerated and flat skin surfaces. Tumors exhibiting these traits were significantly more likely to exhibit higher aggressiveness levels.
The growing interest in artificial intelligence within skin cancer care is evident, yet Sam Polesie noted its limited practical applicability in healthcare settings to date. He stressed the importance of defining clear application areas where AI research could yield tangible benefits for Swedish healthcare.
Polesie expressed optimism, stating, “We believe that one such application area could be the preoperative assessment of suspected skin cancers, where more nuanced conclusions can influence decisions. The model we’ve developed needs further refinement and testing, but the way forward is clear – AI should be integrated where it actually adds value to decision-making processes within healthcare.”
The study was conducted using images collected from dermatological healthcare at Sahlgrenska University Hospital between 2015 and 2023, bolstering the significance of this research in the ongoing evolution of dermatology practices.