Version française

A new ‘deep learning’ technique

September 26, 2018 |
Visual analysis of histological cross-sectional images and their annotation by anatomopathologists to define regions of interest for future analyses require considerable time commitment. The Center for Microscopy and Molecular Imaging (CMMI) has now developed a method to automate part of this process.

Researchers used to manually demarcate hundreds, or even thousands, of regions of interest in images from which relevant information could be extracted. Using learning algorithms, and especially deep learning, computers can now learn to recognise objects, structures, or shapes based on examples. ‘These computer programs learn the specific characteristics of each object on their own,’ explains Yves-Rémi Van Eycke, researcher and computer scientist at the CMMI's Digital Pathology centre (DIAPath). ‘This lets us automatically detect objects in images.’

Thousands of auto-generated examples

Yves-Rémi Van Eycke has worked on glandular epithelium in colorectal tissue, with a view to teaching the programme to reliably identify glands in a histological section. After asking an expert to circle these structures in a hundred photos, the computer scientist also created an algorithm that used these actual examples to automatically generate hundreds of thousands of virtual examples needed for learning. ‘Automatically generating examples considerably reduces the human effort necessary,’ he continues. ‘What used to take months can now be done in less than a day.’ (1)

(Almost) ignore colour and form variations

‘Our method is also more accurate and more reliable with regard to variations in colour and shape, which can hinder computer-based recognition.’ Colouring techniques are used to differentiate proteins and/or cellular components (nucleus, cytoplasm, etc.). However, the proteins highlighted this way and their distribution can vary from one histological section to the next. ‘Different machines, techniques, and manipulations may also result in colour variations,’ adds Van Eycke. ‘In addition, the glands in healthy colorectal tissue have a regular shape and are made up of a string of fairly homogenous cells. In cancer tissue, however, the glandular epithelium can be highly deformed. Our new method can, to a certain degree, see beyond these variations.’

What about other types of glands?

This new deep learning method is highly promising for cancer researchers and anatomopathologists. And not only for work on colorectal tissue: ‘We now have a new tool that can be adapted to detect glandular structures,’ says Van Eycke. ‘We have tested it with prostate glands, and we could improve the program's performance by adding more specific examples, but it has “learned” the basic concept.’

Notes:
(1) Van Eycke YR et al., Segmentation of glandular epithelium in colorectal tumours to automatically compartmentalise IHC biomarker quantification: A deep learning approach in ‘Medical Image Analysis’, July 2018.

Candice Leblanc