AI Model Accurately Diagnoses Skin Cancer
- A new study published in Jaad International demonstrates that a convolutional neural network can assess the differentiation levels of cutaneous squamous cell carcinomas with accuracy comparable to...
- Published September 25, 2024, the research highlights the potential of artificial intelligence to augment clinical expertise in dermatology.
- Cutaneous squamous cell carcinoma (cSCC) is a common type of skin cancer arising from mutations in the cells of the skin's outer layer.
Machine Learning Matches Dermatologists in assessing Skin Cancer Severity
Table of Contents
A new study published in Jaad International demonstrates that a convolutional neural network can assess the differentiation levels of cutaneous squamous cell carcinomas with accuracy comparable to experienced dermatologists. this technology holds promise for improving preoperative decision-making and potentially guiding less invasive treatment options.
Published September 25, 2024, the research highlights the potential of artificial intelligence to augment clinical expertise in dermatology.
Understanding Cutaneous Squamous Cell Carcinoma
Cutaneous squamous cell carcinoma (cSCC) is a common type of skin cancer arising from mutations in the cells of the skin’s outer layer. it’s strongly linked to cumulative ultraviolet (UV) radiation exposure, frequently enough developing on sun-exposed skin exhibiting signs of damage like rough, scaly patches, uneven pigmentation, and reduced elasticity.
Sam Polesie, MD, PhD, Associate Professor of Dermatology and Venereology at the University of Gothenburg and Practicing Dermatologist at Sahlgrenska University Hospital in Gothenburg, Sweden, led the study. He emphasized the cancer’s connection to long-term sun exposure.
Study Design and Methodology
Traditionally, preoperative punch biopsies aren’t routinely performed for suspected cSCC. Instead,specimens are sent for histopathological analysis after surgical excision. This study explored whether a machine learning model could predict the differentiation level of cSCC based on clinical images alone.
Researchers trained a de novo convolutional neural network using 1,829 clinical close-up images of cSCC.The dataset was divided into training (n = 1,329), validation (n = 200), and test sets (n = 300). 68.6% of the cases were well-differentiated. The model’s performance was then compared to assessments from seven independent dermatologists, who also indicated their confidence levels and noted specific clinical features present in each tumor.
Key Findings: AI Performance and Clinical Correlations
The convolutional neural network achieved an area under the curve (AUC) of 0.69 (95% CI = 0.63-0.76). The dermatologists’ combined assessment yielded an AUC of 0.70 (95% CI = 0.64-0.76; P =.79). This demonstrates moderate agreement between the AI model and experienced clinicians, underscoring the inherent complexity of assessing cSCC differentiation.
Specific clinical features were found to correlate with tumor differentiation:
- Ulceration: Moderately or poorly differentiated tumors were 2.34 times more likely to be ulcerated (OR = 2.34; 95% CI = 1.16-4.72).
- Flat Surface Topography: Moderately or poorly differentiated tumors were nearly three times more likely to have a flat surface (OR = 2.94; 95% CI = 1.23-7.01).
Implications for Clinical Practice
The study authors believe this machine learning approach holds notable promise for assisting dermatologists in preoperative decision-making. It could help determine appropriate surgical margins or identify patients who might benefit from less invasive treatments.
However, the model requires further refinement and validation before it can be reliably integrated into clinical practice.