Aug. 22, 2020
Over the past decade, computers using deep neural networks have been able to approach—if not exceed—human performance at object recognition tasks. In 2015, the PReLU-Net deep network became the first computer model to surpass human accuracy on the ImageNet 2012 dataset, with 4.94 percent error compared with humans’ 5.1 percent.
However, object recognition for medical imaging is significantly more challenging than learning to distinguish between images of cats and dogs. This is for several reasons: medical images are much higher-resolution, they contain a wide variety of possible objects, and they often require analyzing multiple features to make an accurate diagnoses.
Despite this greater complexity, recent years have seen several noteworthy breakthroughs with object recognition for medical imaging. In this article, we’ll go over how object recognition is used in medicine and some of the field’s latest developments.
Computer vision and deep learning have a vast range of possible applications, from self-driving cars to detecting manufacturing errors. In medicine, object recognition models have the potential to make diagnoses and detections substantially faster and more accurate.
In 2016, for example, a research team at Beth Israel Deaconess Medical Center and Harvard Medical School developed a highly advanced AI system to interpret pathology images. The team’s deep learning-based model is able to correctly distinguish between cancerous and non-cancerous lymph node cells 92 percent of the time, which is very close to the 96 percent accuracy of a human pathologist. What’s more, when the human and AI analyses are combined, this accuracy improves to a near-perfect 99.5 percent.
An estimated 100,000 Americans die or are permanently disabled every year due to physicians’ diagnostic errors. By assisting with medical imaging diagnoses (including flagging borderline cases for human review), AI models can help reduce these errors and remove some of the load from doctors’ notoriously busy and lengthy work schedules.
Several factors and trends have combined, and will continue to combine, to make AI object recognition a highly promising and viable prospect in medicine:
Object recognition in medicine is carried out through deep neural networks, which use multiple interconnected layers of artificial “neurons” loosely resembling the human brain. Initial layers in the network identify simple patterns such as vertical, horizontal, and diagonal lines. The output of these first layers is then fed into later layers, which are capable of describing arbitrarily complex patterns (e.g. edges, curves, and shapes).
One of the most exciting recent breakthroughs in object recognition for medicine is the use of deep neural networks to detect COVID-19 cases in X-ray images of patients’ lungs. The model is able to achieve an accuracy of over 98 percent for binary classification (i.e. COVID case vs. no case), and an accuracy of 87 percent for multi-class classification (i.e. COVID case vs. no case vs. pneumonia).
Using X-ray images to diagnose COVID-19 offers a viable alternative to PCR tests (swabbing), which have had several issues during the COVID-19 pandemic: lack of tests, high test cost, and potentially long waiting times (on the order of days or weeks). In addition, the study’s authors note that their COVID-19 deep learning model can be used in two ways: either to validate human radiologist screenings, or to provide immediate diagnoses for patients by running in the cloud.
Besides the COVID-19 diagnoses, the sky seems to be the limit when it comes to using deep learning for object recognition in medicine. Below are some of the recent applications of object recognition for medical diagnoses from the past year alone:
While it’s less than a decade old, object recognition in medicine has already seen many promising results, including the potential to diagnose COVID-19 cases via X-ray imaging. As automated object recognition, computer vision and deep learning techniques see growing adoption in the healthcare industry, the Chooch healthcare AI platform is here to help.
Chooch AI unites data collection, model development, training, and deployment on one cost effective AI platform, making it easier than ever for healthcare professionals to get fast, accurate results. Our healthcare customers have used Chooch AI for everything from medical imaging analysis to automated surgical logs and mask and handwashing detection. Chooch AI also is part of the Clara Guardian Healthcare Ecosystem from NVIDIA.
Want to find out how Chooch AI’s computer vision solutions can make your healthcare workflows more powerful and efficient? Get in touch with our team today for a chat about your needs and objectives.