Background: Hypertrophic scars and keloids tend to cause serious functional and cosmetic impediments to patients. As these scars are not life threatening, many patients do not seek proper treatment. Thus, educating physicians and patients regarding these scars is important. The authors aimed to develop an algorithm for a scar screening system and compare the accuracy of the system with that of physicians. This algorithm was designed to involve health care providers and patients.
Methods: Digital images were obtained from Google Images (Google LLC, Mountain View, Calif.), open access repositories, and patients in the authors' hospital. After preprocessing, 3768 images were uploaded to the Google Cloud AutoML Vision platform and labeled with one of the four diagnoses: immature scars, mature scars, hypertrophic scars, and keloid. A consensus label for each image was compared with the label provided by physicians.
Results: For all diagnoses, the average precision (positive predictive value) of the algorithm was 80.7 percent, the average recall (sensitivity) was 71 percent, and the area under the curve was 0.846. The algorithm afforded 77 correct diagnoses with an accuracy of 77 percent. Conversely, the average physician accuracy was 68.7 percent. The Cohen kappa coefficient of the algorithm was 0.69, while that of the physicians was 0.59.
Conclusions: The authors developed a computer vision algorithm that can diagnose four scar types using automated machine learning. Future iterations of this algorithm, with more comprehensive accuracy, can be embedded in telehealth and digital imaging platforms used by patients and primary doctors. The scar screening system with machine learning may be a valuable support tool for physicians and patients.
Clinical question/level of evidence: Diagnostic, II.
Copyright © 2022 by the American Society of Plastic Surgeons.