Machine Vision: How Can It Help Pathologists and Rheumatologists?

Machine vision (sometimes called robotic image analysis) can be seen as a threat to pathologists and radiologists; Steven Guan (PhD, of MITRE Corporation in McLean, Virginia) and his colleagues disagree. The research team explains that using new technologies can reduce the burden on pathologists and other doctors. The primary task in the pathology laboratory is to find and characterize the main details of a histological specimen. Automating these processes will help the pathologists select the most important areas of the sample and make the entire process more cost-effective for clinics. Guan and colleagues note that the Hospital for Special Surgery in New York City alone, which provided the material for the study, performs more than 5,000 arthroplasties per year.

Article image

Scientists say that AI-assisted image analysis can only be as good as a pathologist's conclusion. The researchers explored the possibility of using synovial samples from people with rheumatoid arthritis and presented encouraging results from computer-assisted quantification of joint inflammation.


Guan and colleagues note that the sensitivity and specificity of the algorithm they created correlated with the pathologist's scores: 97% and 100%, respectively, from 170 samples taken from patients with rheumatoid arthritis.

Results obtained using the algorithm also correlate with other disease symptoms, such as C-reactive protein, rheumatoid factor, antibodies to cyclized citrullinated bun, and erythrocyte sedimentation rate, reports ACR Open Rheumatology.

The task of the developed algorithm is relatively straightforward: to count the nuclei in synovial tissue samples. On the one hand, this is a fairly simple quantitative measurement of inflammation. On the other hand, there can be hundreds of thousands of nuclei in one sample, which significantly slows down and makes it difficult to count them without machine vision.

The entire synovial membrane was available for analysis, as samples were taken from people with rheumatoid arthritis who underwent an arthroplasty procedure. Tissues were routinely stained with hematoxylin and eosin (H&E), mounted on glass slides, photographed, and subjected to machine vision analysis.

A team of experienced pathologists was also assembled, who double-checked the accuracy of the nuclei count by the algorithm on 10 images. The specialists also evaluated the samples histologically, paying attention to the most inflamed areas. Moreover, gene expression and laboratory analyzes of patients were studied.


On average, the algorithm found about 112 thousand cores in the sample; however, the concentration of H&E did not affect the algorithm's results, as noted by Guan and colleagues.

Another goal of the study was to confirm that nuclear density can be a useful indicator, since it is highly correlated with other, more traditional, test results (e.g. histologic score, gene expression, lab values). However, while the density of nuclei correlates quite strongly with some indicators, no such correlation was found with others. These indicators include the number of affected joints, pain and self-assessment of the condition by the patient and the duration of the disease.

Such an interesting result is not clear, but the researchers suggest that the reason for this outcome is that the samples were taken from patients with arthroplasty, i.e. people with end-stage disease. Accordingly, these results cannot be interpreted for the entire population of patients with rheumatoid arthritis.

Guan and his colleagues noted that the algorithm needs to be refined and is not ready yet for widespread clinical use, but this first step could be the beginning of developing a reliable product that will be successfully used in the daily work of doctors.

Another valid technique in order to study rheumatoid arthritis is to use X-Ray images with and without contrast, as provided by Medical Data Cloud. In the specific case, each image is accompanied by a detailed radiological description. The company selects images having different levels of severity in organ damage. Choosing Medical Data Cloud as a data source means relying on a large and clean dataset in order to train a good AI algorithm with success.