Medical Image Recognition

Automatic detection of pathological changes in the lungs and prioritizing them
iconData Analytics
iconData Science
iconMachine Learning
iconComputer Vision

Visualization of pathological processes plays a crucial role in establishing most clinical diagnoses. However, sometimes the results could be controversial. Various factors influence it: the quality of the equipment, lack of medical skills and knowledge, limited time, etc. One of our clients decided to develop an algorithm for a medical image segmentation solution to distinguish between normal and abnormal chest X-rays (CXR) images. Read on to find out details.


The imaging process significantly impacts the understanding of the disease progression and helps to administer effective treatment much faster. However, most general-purpose algorithms detect any abnormality that CXRs contain. That's why it is challenging to develop a classifier, as there are lots of abnormal findings on CXRs. Meanwhile, we developed a DL model that precisely detects general abnormalities, unseen instances of tuberculosis, and COVID-19. The idea was to use computer vision, statistical and machine learning algorithms.

  • We used computer vision and machine learning algorithms for image processing and recognition. Also, we applied a deep learning system based on EfficientNet-B7 architecture, previously on Image-Net. Our system was trained on CXR images, and each labeled accordingly using the NLP approach.
  • It was obvious that we would face the problem of standardization. However, we had a solution — SciForce medical professionals analyzed and validated the algorithm's results.
  • Later, we checked the system generalizing to the new population, comparing its performance on two separate datasets. Both datasets (the first one provided by our client, and another, ChestX-ray14 was publicly available) consisted of various abnormalities. Our team labeled these datasets. Since potential use cases imply using our solution in novel settings related to other diseases, we used de-identified datasets to evaluate the generalizability to the new populations.
  • Moreover, we received the feedback from doctors on the nature and the importance of errors, which also improved the system's performance from a medical and business perspective.
  • Finally, we simulated the “prioritizing” function: the algorithm was ordering abnormal CXR cases ahead of normal ones.

We provided a solution to detect abnormal chest X-rays. This algorithm automatically detects any pathological changes in the lungs (general abnormalities, unseen instances of tuberculosis, and COVID-19) and prioritizes them. Our solution comes with the code documentation, scientific report and analytics of the resultant data.

Tech Stack
  • Domain Expertise in Healthcare
  • ML/DL
  • Computer Vision
  • Data Science and Data Analytics
  • Programming languages: Python

So this is how transforming CXRs data into a human-readable format would later boost medical professionals' estimation.

Our contacts