Artificial Intelligence in Skin Cancer: Man vs Machine Redux

Share this content:
The use of AI in the diagnosis of skin cancer has real potential, yet current algorithms lack standardization and regulatory approval.
The use of AI in the diagnosis of skin cancer has real potential, yet current algorithms lack standardization and regulatory approval.

In the 2018 study published in the Annals of Oncology, Haenssle and colleagues proved the superior performance of CNN when compared with the detection of melanomas by 58 dermatologists across different countries. The CNN mean sensitivity was 88.9%, and reached a specificity of 82.5% (vs a mean of 71.3% for dermatologists, P < .01), demonstrating that dermatologists could be assisted by CNN image classification.7

The application of CNNs in nonpigmented, nonmelanocytic cases presents a challenge and has not been studied. These cases present with a range of differential diagnoses, from benign to malignant disease, and their images are subject to fine variations.8

Continue Reading Below

To test the efficacy of CNNs to identify nonpigmented lesions, Philipp Tschandl, MD, PhD, the Medical University of Vienna, Austria, and colleagues used a combined CNN (cCNN) composed of dermoscopic and clinical close-up images, pretrained with 14,000 images and their corresponding 2072 real images of skin diseases. CNN test results were compared with that of dermatologists of various expertise levels.8 

In this experiment, cCNN was found to have a significantly higher area under the curve than human raters, and identified lesions better than beginner- and intermediate-level raters—but not significantly better than expert raters.

cCNN sensitivity was significantly higher for beginner raters, and higher in specific diagnoses than beginner and intermediate dermatologists, but not when compared with the lesion-calling of expert dermatologists. The cCNN achieved a higher sensitivity in delivering diagnoses of malignancy such as squamous cell carcinoma and basal cell carcinoma, but performed poorly in benign cases and in rare or atypical cases, as those tended to be underrepresented in the training set, with a low number of images.

Speaking to Cancer Therapy Advisor, Dr Tschandl explained that CNNs are not ready for clinical application, as they don't include all diagnostic classes, have not been tested prospectively, and are subject to verification bias. However, even though CNNs will not likely replace physicians, they can help speed diagnosis, prioritize cases in low-access populations, and educate those who are not yet diagnostic experts.

Dr Tschandl also emphasized that CNN classifiers, having failed to show clinical benefit, are not approved by the US Food and Drug Administration (FDA) or the European Medical Association (EMA).

Available CNNs are also unable to distinguish atypical melanomas and nonpigmented cases, which constitute the majority of skin cancer cases. Furthermore, the system lacks the accountability typically applicable to physicians in cases of misdiagnoses and malpractice.                

Importantly, false-positive results in CNNs are a real concern, especially from a cost perspective. For instance, a recent skin-surveillance program was found to increase costs.3,9 Autonomous skin surveillance conducted by classifiers, says Dr Tschandl, is considered a high-risk method.

Page 2 of 3

Related Resources

You must be a registered member of Cancer Therapy Advisor to post a comment.

Sign Up for Free e-newsletters



Regimen and Drug Listings

GET FULL LISTINGS OF TREATMENT Regimens and Drug INFORMATION

Bone Cancer Regimens Drugs
Brain Cancer Regimens Drugs
Breast Cancer Regimens Drugs
Endocrine Cancer Regimens Drugs
Gastrointestinal Cancer Regimens Drugs
Gynecologic Cancer Regimens Drugs
Head and Neck Cancer Regimens Drugs
Hematologic Cancer Regimens Drugs
Lung Cancer Regimens Drugs
Other Cancers Regimens
Prostate Cancer Regimens Drugs
Rare Cancers Regimens
Renal Cell Carcinoma Regimens Drugs
Skin Cancer Regimens Drugs
Urologic Cancers Regimens Drugs