“AI is meant to be an enhancement strategy, a support tool to improve our diagnostic abilities,” Dr. Patel, a Mohs surgeon who is director of cutaneous oncology at the George Washington University Cancer Center, Washington, said during the ODAC Dermatology, Aesthetic & Surgical Conference. “Dermatologists should embrace AI and drive how it is utilized – be the captain of the plane (technology) and the passenger (patient). If we’re not in the forefront of the plane, we’re not to be able to dictate which way we are going with this.”
Dr. Vishal A. Patel
In 2019, a group of German researchers found that AI can improve accuracy and efficiency of specialists in classifying skin cancer based on dermoscopic images. “I really do believe this is going to be the future,” said Dr. Patel, who was not involved with the study. “Current research involves using supervised learning on known outcomes to determine inputs to predict them. In dermatology, think of identifying melanoma from clinical or dermoscopic images or predicting metastasis risk from digitized pathology slides.”
However, there are currently no universal guidelines on how large an AI dataset needs to be to yield accurate results. In the dermatology literature, most AI datasets range between 600 and 14,000 examples, Dr. Patel said, with a large study-specific variation in performance. “Misleading results can result from unanticipated training errors,” he said.
“The AI network may learn its intended task or an unrelated situational cue. For example, you can use great images to predict melanoma, but you may have an unintended poor outcome related to images that have, say, a ruler inside of them clustered within the melanoma diagnoses.” And unbeknown to the system’s developer, “the algorithm picks up that the ruler is predictive of an image being a melanoma and not the pigmented lesion itself.” In other words, the algorithm is only as good as the dataset being used, he said. “This is the key element, to ask what the dataset is that’s training the tool that you may one day use.”
Convolutional neural network
In 2017, a seminal study published in Nature showed that for classification of melanoma and epidermal lesions, a type of AI used in image processing known as a convolutional neural network (CNN) was on par with dermatologists and outperformed the average. For epidermal lesions, the network was one standard deviation higher above the average for dermatologists, while for melanocytic lesions, the network was just below one standard deviation above the average of the dermatologists. A CNN “clearly can perform well because it works on a different level than how our brains work,” Dr. Patel said.
In a separate study, a CNN trained to recognize melanoma in dermoscopic images was compared to 58 international dermatologists with varying levels of dermoscopy experience; 29% were “beginners,” with less than 2 years of experience; 19% were “skilled,” with 2-5 years of experience; and 52% were “experts,” with at least 5 years of experience. The analysis consisted of two experiments: In level I, dermatologists classified lesions based on dermoscopy only. In level II, dermatologists were provided dermoscopy, clinical images, and additional clinical information, while the CNN was trained on images only. The researchers found that most dermatologists were outperformed by the CNN. “Physicians of all different levels of training and experience may benefit from assistance by a CNN’s image classification,” they concluded.