IxorThink makes its first successful steps into the world of digital pathology
Digital pathology
A pathologist analyses and detects diseases (e.g. cancer) based upon the microscopic analysis of very small tissue samples. Tiny slices of tissue are colored and fixed upon glass slides. The pathologist looks through a microscope to analyze the structure and cell deformations to determine the disease. Since a few years, it is possible to make digital images of those glass slides using whole slide scanners. The field of viewing and sharing those digital images on a computer screen is known as digital pathology, analyzing them is known as computational pathology.
The start of the journey into computational pathology
During a family dinner, an impassionate discussion between medically trained family members and engineers arose: Can a computer perform the task of a highly trained pathologist? Is it possible to design and train a model that distinguishes abnormal from normal tissues given digital images of glass slides?
So IxorThink started its journey into digital pathology.
First of all, we started our search for pathologists who use whole slide scanners. Next, we needed to convince them of the possibilities of AI.
At St Jan Brugge, the pathological anatomy lab was the first in Belgium to go digital. Dr. De Paepe, Anatomist Pathologist at AZ St Jan Brugge, was intrigued by our project. Together we created a pilot study, approved by the ethical committee of AZ St Jan: “Image recognition techniques on digital images of colon and gastric biopsies.”
Pilotstudy
The aim of the study was to build an algorithm that would predict the chance of a tissue being abnormal. So the digital images were classified into two groups: normal and abnormal tissue. Dr. De Paepe collected digital images of colon and gastric biopsies.
IxorThink built and trained several algorithms. Afterwards, the performance of the different algorithms was tested on unseen images.
Results
The best performing algorithm achieved an accuracy of 95% on unseen images, with 98% precision and 99% recall on abnormal tissue. The algorithm achieved 88% precision and 89% recall on normal tissue.
Our pilot study is accepted for poster presentation at the Digital Pathology Euroscicon Congress in Madrid this August. At this time, the results and in-depth description of the study will be made public.
Conclusion
When good quality data and correct labeling is available, computer algorithms will be capable of performing a specific analysis of digital images of glass slides, equal to the analyses performed by pathologists. But for now, every specific analysis task needs a new algorithm trained on a data set put together for that specific task. Collecting all that data and the specialist labeling remains the biggest hurdle to implementing computational pathology in the daily workflow of the pathologist.
