Artificial intelligence

Artificial intelligence (AI) uses mathematics and computer science methods, allowing one to assign the equivalent of a function as an outcome of an initially disorganized set of data collected from real life. While the initial concept dates back to 1950 (Turing), the development of efficient algorithms and the demonstration of their validity continued on silently for many decades.

We have seen groundbreaking applications of AI in the field of medicine in recent years, particularly in melanoma recognition and image interpretation. Other types of use can also be found in care organisation and clinical trials, and the use of therapeutics will soon be influenced as well. In fact, in recent years the FDA has carefully taken into consideration the use of AI in medicine, working towards the definition of appropriate regulations, with nearly 120 algorithms approved so far.

While there are still some challenges in terms of methodology, medical education and sociology to make all stakeholders comfortable with the technology, the implementation of AI is within our grasp and will benefit physicians and patients by supporting or enabling the efficiency and consistency of diagnoses as well as saving time.

Along with the perspectives, challenges and benefits of AI in medicine, two specific projects will be presented during the EADV Spring Symposium.

The first one is DermAI, which aims to improve the existing Teledermatology processes between Primary Care Units and Hospital Dermatology Departments in the Portuguese National Health System (NHS) for skin lesion diagnosis through the use of AI.
The proposed framework will change processes by assisting both general practitioners and dermatologists: the first will benefit by a computer vision-based mobile application to help on the standardization of image acquisition in dermatology, while the second will help in case prioritisation through an AI-powered risk prioritisation and decision support platform based on retrospective data.

The second project presents deep learning neural network for melanoma diagnosis. Recently, the medical imaging field drew its attention to convolutional neural network (CNNs), deep learning (DL) models that function in a similar way to the neurons connections of the human visual cortex. Based on the linear operation “convolution” in multiple processing layers, a CNN structure includes an input layer processing data, a series of hidden layers extracting the features and an output layer. As the number of hidden layers grows, CNNs become “deep” (DCNN): the higher the number of layers of the DCNN, the higher the level of abstraction reached, the larger is the set of supervised data needed for adequate training. In the last few years, DCNNs have grown in importance in decision support systems for dermoscopic analysis of metastatic melanoma and skin lesions in general, as well as various CNNs models constructed in international challenges on large datasets.