AI: State-Of-The-Art and Commercialisation Status in Diagnostics
Artificial intelligence (AI) is revolutionizing medical diagnostics. The state-of-the-art results have already demonstrated that software can achieve fast and accurate image-based diagnostics on various conditions affecting the skin, eye, ear, lung, breast, and so on. These technological advancements can help automate the diagnosis and triage processes, accelerating the process to speed up the referral process especially in urgent cases, freeing up expert resources, offering the best accuracy everywhere regardless of skill levels, and making the processes more widely available. This is a ground-breaking development with far-reaching consequences. Naturally, many innovators are scrambling to capitalize on these advancements.
Our report Digital Health & Artificial Intelligence 2020: Trends, Opportunities, and Outlook has examined this trend. This report considers the trend towards digital and AI applications in health. It outlines the state-of-the-art in AI-based diagnosis of various conditions affecting the skin, eye, heart, breast, brain, lung, blood, genetic disorders and so on. The data sources employed are diverse including dermoscopic images, fundus images, OCT, CT, CTA, echocardiograms, electrocardiogram, mammography, pathology slides, low-res mobile phone pictures and more. This report then identifies and highlights companies seeking to capitalize on these technology advances to automate the diagnostic and triage process.
Furthermore, this report considers the trend of digital health more generally. It provides a detailed overview of the ecosystem and offers insights into the key trends, opportunities and outlooks for all aspects of digital health, including: Telehealth and telemedicine, Remote patient monitoring, Digital therapeutics / digiceuticals / software as a medical device, Diabetes management, Consumer genetic testing, Smart home as a carer and AI in diagnostics.
Significant funding is flowing to start-ups and R&D teams of large corporations who develop AI tools to accelerate and/or improve the detection and classification of various diseases based on numerous data sources ranging from RGB images to CT scans, ECG signals, mammograms and to pathological slides. The state-of-the-art results demonstrate that software can do these tasks faster, cheaper, and often more accurately than trained experts and professionals.
This is an important development which, if successful, can have far-reaching consequences: it can make diagnostics much more widely available and it can free up medical experts' time to focus on more complex tasks which currently sit beyond the capabilities of AI-based automation. The technology is today making leaps forward, but technology is only a piece of the puzzle, and many other challenges will need to be overcome before such software tools are widely adopted. However, the direction of travel is clear.
This trend is today on the rise because (a) the availability of digitized medical data sources is rapidly increasing, offering excellent algorithm training feedstock, and (b) advancements in AI algorithms specially trained deep neural networking are enabling software to tackle tasks which it hitherto could not do.
Our report Digital Health & Artificial Intelligence 2020: Trends, Opportunities, and Outlook outlines many such advancements and identifies some of the key companies pursuing each opportunity. In the remainder of this article, we briefly outline two specific cases: eye disease and skin disease.
Diabetic retinopathy is a complication that affects the eye. Researchers from India have recently shown that the software accurately interprets retinal fundus photographs to enable a large-scale screening program to detect diabetic retinopathy. The software is trained to make multiple binary classifications, allocating a risk level to each patient. The algorithm was trained and tuned on a total of more than 140k images. The machine matched and exceeded the sensitivity and selectivity level achieved by trained manual experts. The software achieved 92.1% and 95.2% sensitivity and selectivity, respectively.
Naturally, there is a strong business case here, and many are seeking to capitalize on it. One example is IDx, based out of Iowa in the US, who has designed and developed an algorithm to detect diabetic retinopathy. Their AI system achieves a sensitivity and specificity of 87% and 90%, respectively. In as early as 2017, it was tested at 10 sites across the US on 900 patients.
A very insightful test in eye clinics is the OCT (optical coherence tomography), which creates high-resolution (5um) 3D maps of the back of the eye and require expert analysis to interpret. OCT is now one of the most common imaging procedures with 5.35 million OCT scans performed in the US Medicare population in 2014 alone. This creates a backlog in processing and triage, and such delays can be harmful when they cause avoidable treatment delay for urgent cases.
DeepMind (Google) has demonstrated an algorithm that can automate the triage process based on 3D OCT image. Their algorithm design has some unique features. It consists of two stages: (1) a segmentation network and (2) a classification network. The first network will output a labelled tissue segmentation map. Based on the segmented maps, the second network will output a diagnosis probability for over 50 eye-threatening eye conditions and provide referral suggestion. The first part was trained on 877 sparely and manually segmented images and the second network on 14,884 training tissue maps with confirmed diagnosis and referral decision. This database is one of the best curated medical eye databases worldwide.
This two-stage design is beneficial in that when the OCT machine or image definition changes, only the first part will need to be retrained. This will help this algorithm become more universally applicable. In an end-to-end training network, the entire network would need to be retrained.
DeepMind demonstrated that performance of their AI in making a referral recommendation, reaches or exceeds that of experts on a range of sight-threatening retinal diseases. The error rate on referral decision is 5.5%, exceeding or matching specialists even when specialists are given fundus images as well as patient notes in addition to the OCT. Furthermore, the AI beat all retina specialists and optometrists on selectivity and sensitivity measures in referring urgent cases. This is clearly the first step, but an important one that truly opens the door.
Researchers at Heidelberg have already demonstrated that trained deep neural networks, in this case based on Google's Inception v4 CNN architecture, can recognize melanoma based on dermoscopy images. These researchers showed that the software achieves 10 percent more specificity than human clinicians when the sensitivity was set at a level matching human clinicians. The machine can achieve a high 95% sensitivity at a 63.8% specificity.
This is a promising result that shows such diagnostics can be automated. Indeed, multiple companies are automating detection of cancer diseases. One example is SkinVision, from the Netherlands, which seeks to offer a risk rating of skin cancer based on relatively low-quality smartphone images. They trained their algorithm on more than 131k images from 31k users in multiple countries. The risk ranking of the training images were annotated by dermatologists. Studies show that the algorithm can score a 95.1% sensitivity in detecting (pre)malignant conditions with 78.3% specificity. These are good results although the specificity may need to improve as it could unnecessarily alarm some patients.
The business cases are not just limited to cancer detection. Haut.AI is an Estonian company that proposes to use images to track skin dynamics and offer recommendations. One example is that their AI can be a simple and accurate predictor of chronological age using just the anonymized images of eye corners. The networks were trained on 8414 anonymized high‐resolution images of eye corners labelled with the correct chronological age. For people within the age range of 20 to 80 in a specific population, the machine reaches a mean absolute error of 2.3 years.
There are naturally many more start-ups active in this field. Some firms are focused on health diagnostic whilst others are seeking to use the AI to create tailored skincare regimes and product recommendation. The path to market, and the regulatory barriers, for each target function will naturally be different.
To learn more about this exciting field, please see our report Digital Health & Artificial Intelligence 2020: Trends, Opportunities, and Outlook by visiting www.IDTechEx.com/digitalhealth. This report outlines the state-of-the-art in the use of AI in diagnosing a range of medical conditions. It also identifies and discusses the progress of various companies seeking to commercialize such technological advances. Furthermore, this report considers the trend of digital health more generally. It provides a detailed overview of the ecosystem and offers insights into the key trends, opportunities and outlooks for all aspects of digital health, including: Telehealth and telemedicine, Remote patient monitoring, Digital therapeutics / digiceuticals / software as a medical device, Diabetes management, Consumer genetic testing, Smart home as a carer and AI in diagnostics.