Will AI replace dermatologists

In this opinion piece, Anmol Arora (University of Cambridge, UK) discusses the current applications of AI in dermatology, the barriers that prevent further implementation and what AI holds for the future of dermatology.

Introduction

AI is conventionally defined as the ability of a computer system to perform tasks that are traditionally thought to require human intelligence, such as deductive reasoning and learning. Dermatology has been heralded as one of the earlier specialties to be affected, given that current AI research focuses on image and data analysis, two key elements of a dermatologist’s role. Types of artificial intelligence, such as convolutional neural networks (CNNs), are modeled based on the animal visual cortex. A triad of advances in computer science, computing power and big data have enabled AI to become embedded in translation softwares, recommendation engines and early autonomous driving systems.

Early studies have noted that in some circumstances – usually in isolated datasets – AI algorithms have been able to outperform dermatologists at diagnosing skin cancer from images. Most of these studies have utilized supervised learning systems, where an AI algorithm is provided with a large dataset of labeled skin images, some of which show cancer and others are normal. There are still barriers to adoption that must be overcome before AI can revolutionise dermatology but there are already early arguments that AI systems may threaten conventional medical training pathways.

Existing research efforts

There are four main AI techniques that have been applied in dermatology:

1. Classification
2. Prediction
3. Augmentation
4. Segmentation

Classification refers to the ability of AI to detect dermatological diagnoses from images . Prediction involves forecasting the diagnosis in advance, based on historical images and data. Segmentation involves identifying borders and objects within an image, which assists dermatologists with interpretation. Finally, augmentation refers to altering the data within a dataset to produce more data for training models. These four techniques are closely intertwined. For example, augmentation is a tool for producing more suitable data for the training of classification models. Similarly, segmentation may be utilized to produce labeled data for training other algorithms.

Dermatology is not the only specialty being impacted by the development of AI. Indeed, lessons can be learned by analyzing patterns of AI propagation in other specialties. For example, ophthalmologic AI systems are currently being utilized in practice to detect diabetic retinopathy from clinical fundus images . However, these systems require supervision, they have limitations and they can only diagnose one condition. When AI systems ultimately integrate with dermatological care pathways, they may follow a similar trajectory. AI systems will require human supervision, at least in the early stages. There will also be a need for dermatologists to confirm the findings of the AI algorithms and to communicate the results to patients.

Barriers to implementation

There are several technical limitations to AI development which will require resolution before systems can be widely implemented, none of which are unique to dermatology. These limitations include:

• Limited availability of publicly available datasets
• Lack of external validation for AI algorithms
• Risk of systematic algorithmic bias if datasets are not diverse

AI algorithms are limited by the data that is utilized to develop them. In recent years, the issue of ‘data poverty’ has arisen as a description of the problem that there are certain patient groups who are not represented in training datasets . If data from those patients are not utilized to train algorithms, then the algorithms will not be able to serve them. There is a risk of algorithmic bias, whereby underrepresented patients are inadvertently discriminated against. Other risks include automation bias, where doctors may become reluctant to overrule the output of AI algorithms and accountability, since it is currently unclear who holds responsibility for cases where the algorithms are incorrect.

Any AI solutions must be acceptable both to patients and clinicians. Indeed, dermatological AI research has been accompanied by the development of patient and clinician groups focused on AI co-design . The Food and Drug Administration (MA, USA) has expressed clear intent to regulate AI, including post-market monitoring . The European Medicines Agency (Amsterdam, the Netherlands) and the UK government have also produced reports highlighting issues surrounding the integration of AI in clinical pathways .

Emerging research

Whilst high levels of accuracy have been achieved in detecting dermatological diagnoses from curated datasets, there is still a weak evidence base for safe integration of AI into clinical practice. This is in part due to concerns about algorithm bias, automation bias and accountability. Human-in-the-loop – or HITL – is a method of developing machine learning that leverages both human and machine intelligence to create algorithms. A human is involved in training, testing and supervising the algorithm from development to deployment. In this way, AI may be developed as a tool to assist dermatologists, rather than as a replacement for them.

The growth of AI research in dermatology has attracted widespread speculation about potential utilizations and the possible impact on dermatologists. It is unlikely that AI will replace dermatologists, however it would not be unreasonable to conclude that dermatologists who utilize AI may replace those who do not.

‍Disclaimers:

Anmol Arora has no disclaimers to declare.

The opinions expressed in this feature are those of the interviewee/author and do not necessarily reflect the views of Future Medicine AI Hub or Future Science Group.