Solution for disease diagnostics
based on AI
Automatic detection of anomalies in X-rays, CT scans and mammograms using image recognition methods
The first step in the path of treatment is to correctly examine the x-ray and notice the anomalies on it.

With the introduction of artificial intelligence technology, the neural network will detect anomalies in the medical images of patients, a similar method can be applied to analyze various types of medical images.
Product development

Fluoroscopy is used to diagnose a wide range of diseases and injuries:
lung damage (pneumonia, cancer), fractures and other bone injuries,
part of the diagnosis of the digestive system and much more.
The field of artificial intelligence has made great progress in recent years, especially
in the use of neural networks to solve various kinds of visual analysis tasks.
Accordingly, the idea arose to apply AI to the field of image recognition, where doctors are also involved in image recognition, namely, the analysis of images and, for starters, x-rays.
As a result, the creation
IT products for the recognition of "anomalies" on X-rays using AI
Solution method - prepared dataset with marking of image areas for the presence of pathology
The task is to determine the presence of various pathologies on x-rays, tomograms and mammograms.

The task for the algorithm is to learn how to encode images in such a way that the encoding of images with pathology is very different from encoding images without pathologies.
To do this, we use neural networks and preliminary markup of images (areas in the image and binary markup)
The Solution: Segmentation and Detection
  • Labeled set of images for training models
    A large amount of data marked up for a specific detection problem is needed, which contains both images with "anomalies" and normal images.
  • Training different types of neural networks to achieve the best result
    In such tasks, we use either segmentation or detection neural networks. Often the best result is achieved with a combination of approaches.
  • Model predicts potentially dangerous areas in new images
    At the output of the model, we have areas with a high probability of containing "anomalies" or a clean image without such zones.
Model and situational risks
AI applications
  • -1-
    Precision is what percentage of those patients identified by the model as having pneumonia actually have pneumonia (and thus what reciprocal percentage of physicians would thus mistakenly treat the wrong disease).
  • -2-
    Recall is what percentage of all patients with pneumonia the model will detect (the inverse of this percentage is how many patients with pneumonia the model will mark as healthy)
  • -3-
    Estimation Accuracy
    Our solution for detecting pneumonia on x-rays achieves precision = 0.82 and recall = 0.94. To understand whether such accuracy is sufficient, it is necessary to conduct additional research among doctors and calculate the accuracy of their decisions.
  • -4-
    Unforeseen cases
    Happen all the time in the practice of using machine learning. These are the test cases that we didn't see during the training. Knowing the area of research, it is necessary to protect yourself as much as possible from such cases, since the model may incorrectly identify them as a disease.
Potential for partial and full implementation of AI
  • "Doctor's Advisor"
    - Second opinion will increase
    diagnostic accuracy
    - Control of risks in case of divergence of opinion between AI and doctor
    - Solving the problem of distrust on the part of society
  • "Doctor Replacement"
    - Solving the problem of lack of staff
    - Quality standardization medical care
    - The disappearance of the human factor
Challenges in AI Implementation
  • 1
    Comparison with the quality of classification performed by a doctor
  • 2
    Systems approach
    Standardization of equipment and images
  • 3
    Constant control of the accuracy of models and standardization of models
  • 4
    Trust from personal and society
We have been developing advanced solutions based on neural networks since 2016
The economic effect of cooperation with us -
tens of thousands of dollars a day
Our contacts
+7 (967) 215-75-05