All news

Celsus experts have become co-authors of a scientific article on the use of AI in the diagnosis of rectal cancer


The scientific and practical journal of the Association of Coloproctologists of Russia published an article on the use of artificial intelligence in the MRI diagnosis of rectal cancer. Experts of the NMIC of Coloproctology named after A.N. Ryzhykh and the National Center for Service Integration worked on it together with specialists of the AI department of Celsus - Nikolai Filatov and Evgeny Nikitin

What was investigated?

Within the study’s framework, the task was to evaluate the possibility of developing a medical decision support system based on AI technologies that would help doctors with MRI diagnosis of rectal cancer. Namely, it determined the localization and segmentation of the primary tumor.

We had 450 MRI studies of patients with rectal cancer (with histological verification) and 450 MRI studies of patients without a tumor were used to develop the model. These data were marked up by doctors of the NMIC of Coloproctology named after A.N. Ryzhykh and then used in the construction of an artificial intelligence model.

Why is this important?

Colorectal cancer is the second most common type of malignant neoplasms after lung cancer. More than 600 thousand new cases of the disease are detected annually in the world. In Russia, this figure is 50 thousand per year, and not all cases are detected even at a late stage (no more than 70%).

As with many other forms of cancer, the probability of cure or long-term remission in rectal cancer directly depends on the stage at which the disease was detected. If it is detected at the first two stages, the effectiveness of treatment can be quite high, but with late diagnosis, the mortality rate increases many times and can be more than 40%. The situation is further aggravated by the fact that in the early stages of the disease does not always manifest itself (the patient may not have any symptoms at all).

That is why it is so important to detect colorectal cancer as early as possible, automating the process of its diagnosis and creating additional self-testing tools for the diagnostician – including using artificial intelligence technologies.

What did we find out?

In short: the artificial intelligence model developed as part of the study showed good accuracy (77.0%), excellent sensitivity (98.1%), but unsatisfactory specificity (45.1%).

What does it mean? The model finds and visualizes the primary tumor well, determines its localization, but at this stage of its development gives too many false positive results in healthy patients. To improve the model, which would make it applicable in screening activities, further experiments with training parameters and an increase in the amount of data for subsequent training of the model are necessary.

Learn more about the progress and results of the study!

We thank the experts of the NMIC of Coloproctology named after A.N. Ryzhykh and the National Center for Service Integration for the opportunity to work together and fruitful cooperation. Such studies are important and necessary – and, hopefully, there will be more of them!