An AI-Based Skin Analysis App

Zahra Jazini

December 6, 2022

How Is AI Used In Dermatology?

Artificial intelligence (AI) has emerged as the major research frontier in computer science.
Although AI has been used in many scientific fields, its use in dermatology is relatively new and limited. Dermatologists must have substantial knowledge of AI concepts because skin conditions, with their plentiful medical and dermatoscopic data and images, have the prospect of being the next large subject in the use of artificial intelligence in medicine.
Studies in artificial intelligence have already been conducted on skin disorders such as skin cancer, psoriasis, dermatitis, and candidiasis.
This article provides an overview of artificial intelligence (AI) and innovations pertinent to anesthesiology, evaluating both existing apps and future potential.

Terminologies of Artificial Intelligence

AI and machine learning strength over the last two decades as a result of improved hardware and software technologies have led us to face the possibility of AI to keep improving present healthcare methodologies.
So, AI research is already ongoing in several medical fields, including dermatology.
However, dermatological AI is still in its early stages compared to some fascinating technical AI applications such as radiology.
With the changing scenario and increased research in dermatological AI, one can anticipate that the use of artificial intelligence in dermatology will tremendously reduce the gap between doctors at various levels of medical facilities and improve diagnostic accuracy.

The Current State of Artificial Intelligence in Dermatology

AI has gradually gained relevance in various fields of dermatology, such as skin cancer, skin conditions, and psoriasis, throughout the last ten years.

AI application in skin cancer

Researchers have been investigating the use of artificial intelligence (AI) to optimize or boost the latest testing procedures for skin cancer. Nasr-Esfahani et al. were the first to train a neural network for melanoma detection, and their proposed method had a sensitivity and specificity of 0.81 and 0.80, respectively. Stanford University published a study on deep learning of malignant tumours in 2017. They demonstrated skin lesion categorization using a single convolutional neural network. It was trained end-to-end from images using only pixel resolution and illnesses labels as inputs. A convolutional network was trained using 129,450 clinical images representing 2,032 distinct diseases. The effectiveness of the system was evaluated in comparison to 21 board-certified dermatologists using biopsy-verified clinical images and two crucial binary classifications of cases: keratinocyte squamous cell carcinoma versus benign seborrheic keratoses and malignant melanomas versus benign nevi. The first case involved identifying the most common cancers, while the second involved identifying the deadliest skin cancer. It was discovered that the machine could identify and classify skin cancer with the same accuracy as board-certified dermatologists. This was a pivotal publication on the use of AI in dermatology. However, because they also do not include demographic information, the validation of their work is questionable.

Although it was widely acknowledged applying deep-learning technologies to skin cancer, could improve the high specificity of skin cancer screening too. It was assumed the number of training images required for such a practice could become enormous.

AI can also be used to diagnose skin cancer at the histopathological level. Hekler et al. looked at 695 lesions that were classified by an expert histopathologist according to current guidelines. A convolutional neural network was trained using 595 of the resulting images (CNN). Another 100 H and E section images were used to compare the CNN results to 11 histopathologists. To test for significance (P 0.05), three combined McNemar tests comparing the results of the CNNs test runs in terms of sensitivity, specificity, and accuracy were predefined. Over 11 test runs, the CNN achieved a mean sensitivity/specificity/accuracy of 76%/60%/68%. In comparison, 11 pathologists had a sensitivity/specificity/accuracy of 51.8%/66.5%/59.2%, respectively.

As a result, they concluded that CNN outperformed 11 histopathologists in the classifying of histopathological cancer images, indicating that it has the potential to aid in human melanoma diagnosis.