Hello, Skychain community!
We are ready to share the recent news about Skychain project development. In particular, we will cover the news about our solution on prostate cancer.
Unfortunately, during our internal testing we have stumbled upon a problem of data quality. The second dataset, which was acquired from our partners, had some serious issues in terms of quality of their preparation and requires us to wait for some more time since we need it to be labelled more thoroughly and accurately. But we have some good news about the whole situation — after we initially started training the neural network on the bad quality dataset, the accuracy of it significantly fell. This is the sign that neural network is quite good already and it showed us itself that we are “feeding” it with the bad data.
We are now expecting the finish of dataset relabelling. The dataset is about 90% ready now and we are ready to retrain our dataset on a new, high quality dataset.
Platform for Whole Slide Image labeling
In order to prevent situations like this in the future, we are planning to create our own platform for labeling of digital pathology slides. We aim to build a small team of experienced independent pathologists, who will label the images and will be well rewarded for that.
Using our own platform for image labeling would also allow us to significantly expand the pool of doctors involved in labeling in order to avoid similar mistakes when working with other areas of digital pathology when developing new neural networks of the project.
As part of our labeling platform, we will be able to receive high-quality slides, regardless of third parties and other medical institutions. In order to significantly increase the quality of the prepared labeling, we need to make a two-stage labeling scheme.
At the first stage, doctors will deal directly with the layout of the slides. With the help of tools for highlighting areas, doctors will select specific subject areas pixel by pixel, assigning them certain classes (labels).
Specialists-histologists for the first stage of labeling will require at least 3 years of experience. The labels put down by doctors will be saved on the server.
After that, the slides labeled by doctors will be sent to more experienced specialists (supervisors) for label confirmation. The supervisor will be a specialist with at least 5 years of experience. The supervisor will have to make a decision on each of the labels given by the doctors. The supervisor will have 3 different choices for each specific label:
• “confirm” — after that the final category is assigned to the label;
• “edit” — after that, the supervisor can change the boundaries of this class using the labeling tools.
• “reject” — after that the label is assigned the value of normal tissue, that is, the label put by the doctor at stage 1 is removed.
After the doctor-supervisor has finished editing the slide, he must finally confirm the accuracy of other labels on the slide. After that, the whole labeling made by the supervisor will be saved.
If you are a pathology practitioner and you are interested in helping our project to achieve its goals in your spare time (of course, we will pay you), please write me an email to firstname.lastname@example.org to know more.
As soon as we recieve all the quality data and retrain our neural network, we will finally start testing our solution on the data from various medical institutions. We will let you know as soon as we get first results!
Thanks for the support and stay tuned for future updates!
Alexander Oksanenko, Skychain Team
If you have any questions about Skychain, don’t hesitate to write to Alexander Oksanenko on Telegram and on email: email@example.com.