Preparation for prostate cancer solution testing

Hello, Skychain community!

You haven’t heard anything from Skychain team for a while, but we would like to change it. The COVID-19 outbreak had a minor influence on the project, the team has been on a lockdown for two months now, every team member is safe and is working remotely from home.

As for the project — in this post we would like to share the outline of the tests, which we are about to conduct in the nearest time. This result of it will indicate the state of our solution on prostate cancer, which hopefully will be good enough and will display that it is ready for clinical trials.

The test will be conducted in June and will be held with the help of our two major partners — Moscow Central Hospital and P.A. Hertsen Moscow Oncology Research Center, where Central Hospital will provide us with the validation dataset and Research Center will provide us with the reference standard and specialists.

The whole test will feature two major stages and both of them are quite important for project development. During the first test we will do a competition, comparing the results between two groups of three doctors of 1,3 and 10 years of practical experience and our neural network. At the second test, the doctors will do the diagnostics with the help our neural network, so it could provide them with the initial evaluation of the digital slide. Afterwards, the results of their diagnostics with the help of neural network will be compared to their results without it during the first test. As a result, we will be able to trace how our solution influences the overall performance of doctors, especially the accuracy and speed.

The layout of tests

Before the conduction of the test we will require the reference standard — all the images need to be analyzed by the consensus of three specialists with more than 15 years of experience. Each specialist will provide the diagnosis on each and every of 720 histopathological slides, which form the total data of 60 patients. The conformity of diagnostic parameters of at least two specialists will form the ground truth for each medical image. The diagnostic parameters, which must be provided by each specialist, will include:

  • subtype of cancer (acinar, ductal or foam-cell);
  • Gleason score (the grading system used to determine the aggressiveness of prostate cancer);
  • tumor area percentage (a necessary part of the diagnosis);
  • subjective evaluation of case complexity (the dataset will be separated in two parts, which will be balanced to be equal in their complexity)

After the reference standard is formed, the results from both neural network and doctors will be compared to it.

First stage

During the first stage we will test our neural network in a real competition with 6 pathology specialists of 1, 3 and 10 years of professional experience. In order to make the test clearer, we will have divide them into two groups with the same experience.

The layout of the first stage

Doctors of the same experience (1, 3 and 10) will require to analyze 120 images of 10 patients and provide their diagnosis. The diagnostics will require them to list following cancer indicators:

  • subtype of cancer (acinar, ductal or foam-cell);
  • Gleason score (the grading system used to determine the aggressiveness of prostate cancer);
  • tumor area percentage (a necessary part of the diagnosis).

Skychain neural network will do the same. Afterwards, the relative results will be compared to the reference standard. We will also evaluate various metrics of accuracy, such as Cohen’s kappa and others.

We estimate that at this stage our neural network is capable of showing the results comparable to a specialist with 3 years of experience.

Second stage

The layout of the second stage

At the second stage, the same doctors of the same experience will analyze the other 120 images, but this time they will be assisted by Skychain neural network. Our neural network will provide the instant initial analysis for the doctor, which must be either confirmed and accepted by the doctor or be manually altered in one or several provided parameters, The parameteres will be the same as in the stage 1:

  • subtype of cancer (acinar, ductal or foam-cell);
  • Gleason score (the grading system used to determine the aggressiveness of prostate cancer);
  • tumor area percentage (a necessary part of the diagnosis).

To make the datasets equal in their complexity of cases, the images will be divided into the datasets according to their complexity score set up by the reference standard experts.

Afterwards, the performance of the doctors, including the speed and accuracy of the diagnosis, will be compared to the reference standard, and then to the results of the exact same doctors in the first stage. By doing this, we will see how the neural network influences the performance. We will evaluate the change in both accuracy and speed to prove that doctors with our solution assisting them perform much better than without it.

Please let us know what you think about the framework in our Telegram:

Join Skychain on social media: Twitter, Facebook.

Alexander Oksanenko, Marketing Director

Blockchain infrastructure aimed to host, train and use artificial intelligence (AI) in healthcare. Our website: https://skychain.global/