The automatic identification of cell nuclei in Feulgen-stained histological sections and automated detection of tumour regions in HE-sections forms the basis for a majority of the DoMore! applications. These two tasks are different and treated in separate projects, described in the following.
Automatic Segmentation of Cell Nuclei
The detection and segmentation (draw around) of cell nuclei in Feulgen-stained histological sections is required for the subsequent analyses of DNA ploidy status and Nucleotyping. The Feulgen-stain binds specifically to DNA and allows the analysis of chromatin properties such as amount and structure. Visually the task of identifying and segmenting cell nuclei seems relatively simple to the human eye, but designing robust computer algorithms for this task has been a great challenge for digital image analysis.
In 2012 we published a method for the automated segmentation of cell nuclei in Feulgen-stained histological sections from prostate cancer [ref Nielsen et al., Cytometry A. 2012 Jul;81(7):588-601. doi: 10.1002/cyto.a.22068]. The method worked well in prostate cancer specimens but did not when applied to colorectal cancer specimens, where the cell- and tissue organisation is different. We have used deep learning with convolutional neural networks to develop a novel method for the automatic segmentation of cell nuclei in colorectal cancer. The new method is trained using 300 000 manually delineated cell nuclei in 139 patients and validated in histological sections from 51 independent patients, in which 104 000 manually delineated nuclei were available as a ground-truth in comparison. The image below illustrates result from the neural network model that was adapted to the problem. This new method has proven to work for prostate cancer as well, and a validation study in lung cancer is in progress.
Automatic segmentation of cancerous tumour regions
Pathologists examine tumours and classify them as cancerous or non-cancerous. Our analyses of cancer patients are carried out in the cancerous tumour regions, and a method to identify these regions automatically is thus required. We have used deep learning with convolutional neural networks trained on a dataset of 2573 semi-manually delineated colorectal tumour samples to develop such a method. The resulting computer model is validated on a different set of 857 samples. The approach works well, with 93% sensitivity (the proportion of cancerous tumour region in the ground-truth also identified by the deep learning model) and 97% specificity (the proportion of non-cancerous region in the ground-truth also identified by the deep learning model), i.e. a good correlation with the pathologist’s delineation.
Evaluation of results
To evaluate the quality of our new methods to segment cell nuclei and tumour regions, we need methods that objectively quantify the degree of correspondence with the ground-truth. Ground-truth for the segmentation of cell nuclei is represented by manually segmented cell nuclei, while the ground-truth for cancerous tumour regions is in the form of annotations drawn by a pathologist. There is a lack of consensus for methods to evaluate the quality of these two approaches, although some methods are more frequently used than others, such as the Jaccard index, given two sets X and Y (i.e. the number of overlapping pixels in the predicted segmentation and the ground-truth divided by the total number of unique pixels in the predicted segmentation and the ground-truth). We have implemented a range of measures to evaluate our results robustly.
Oslo, 15th February 2019
Navigate to the other Work Package descriptions:
Download a pdf of the DoMore! project's mid-term report.