Prostate cancer is the leading cause of cancer deaths in men. Diagnosis is based on Gleason grading, which is the most widely used system for determining the severity of prostate cancer from tissue samples. However, Gleason grading is highly subjective with significant variation between experienced pathologists. An improved prostate cancer grading system will both prolong lives and reduce medical costs significantly.
CADESS is a proprietary technology combining consensus-graded tissue data and a new tissue stain with powerful AI and Machine Learning tools: knowledge-based systems, deep learning, and sophisticated classifiers for automatic malignancy grading.
Using biopsy and prostatectomy material with a consensus Gleason grade, we will use pattern recognition and statistical methods to find the features that correlate well with the grade. We will also develop new classifiers for cancer grading that reliably identify malignancy relative to disease end points. By training the classifiers on the disease outcome rather than on the Gleason grade, we aim for better prognostication than currently possible.
We address gland segmentation of tissue stained with PRS-Htx using image analysis methods based on mathematical morphology with adaptive filters. We decompose our images into density maps corresponding to the stroma and the epithelium, and use the stromal density map to determine the exact glandular contours. The method has been successfully tested on more than 5000 glands of different grades, from well-formed glands (Gleason grade 3) to cribriform (grade 4), and single cells (Gleason grade 5).
During 2015 thirteen internationally prominent uropathologists from seven countries remotely graded 650 small images from whole mount sections according to the Gleason system. We observed similar grade variations as seen in other studies; for example, in 60% of the cases more than four pathologists disagreed with the majority. But unlike other studies on intra- and inter-observer grading variation, which are based on entire biopsies or whole mounts, each image in our study contained only one dominant morphological pattern, allowing identification of patterns that cause large grading discrepancies. In September this year the pathologists met in Uppsala to establish a consensus for these patterns.
We developed an automatic method for highly accurate color decomposition of histological images into density maps, one for each stained tissue type. The method decouples intensity from color information by mapping the image data to the Maxwellian Chromaticity plane, and bases the decomposition only on the tissue absorption characteristics of each stain, excluding areas affected by light scattering stains. A quantitative comparison using the root-mean-square-errors (RMSE) of the mixing matrices produced by independent component analysis (ICA), non-negative matrix factorization (NMF), BCD and a ground truth mixing matrix as defined by an experienced pathologist reveals that our method outperforms NMF by between 20% and 40% and ICA at greater margins. Qualitative comparisons reveal a visual difference in the size of the nuclei even for the stronger stain, hematoxylin. The figure shows the results of automatic decomposition of Sirius-Htx+-stained tissue.
Pathologists rely on multiple, contrasting stains to analyze tissue samples, but histological stains are developed for visual analysis and are not always ideal for automatic analysis. We compared thirteen different histological stains from adjacent prostate tissue sections from radical prostatectomies. We evaluated the stains for both supervised and unsupervised classification of stain/tissue combinations. We demonstrated that for a specific tissue type, certain stains perform consistently better than others according to objective error criteria.
The traditional prostate tissue stain, Hematoxylin-Eosin (H&E) dating from 1896, does not allow automatic identification of the prostate glandular structure which is key in malignancy grading. In contrast, PRS stains the connective tissue surrounding the glands red, allowing precise identification of the glandular borders. Both stains use Hematoxylin to stain the nuclear texture. We have also developed a method to add immune-histochemical staining of basal cells to PRS-Htx to discriminate invasive cancer.
The project has been financed by Svenska Vetenskapsrådet (the Swedish Research Council), Hagstrandska Fonden, Vinnova VFT-, Handelsbankens Innovationsstiftelse, and Uppsala University, Department of Immunology, Genetics and Pathology.