Vratis -> Cervical cancer diagnosis

Image processing solutions

Recent projects

Towards rapid cervical cancer diagnosis

The main research objective was the development of a computer-aided system that supports the cervical cancer diagnosis. The core of the system was defined by a combination of different image processing algorithms that analyze images of cytological smears acquired by phase-contrast microscopes.

Schilling T, Miroslaw L, Glab G, Smereka M. Towards rapid cervical cancer diagnosis: automated detection and classification of pathologic cells in phase-contrast images. Int J Gynecol Cancer 2007;17:118–126.

More information can be obtained from lukasz.miroslaw@vratis.com

Phase contrast microscope and cervical cancer

Phase contrast microscopy holds considerable promise in cervical cancer diagnosis, the second most frequent type of cancer worldwide.

Evaluation of samples directly after an examination and large diagnosis spectrum makes this method a recommended alternative to a standard smear test, called a Papanicolau (Pap) test. Although successful in reducing cervical cancer mortality, Pap tests have many drawbacks including high false positives and false negatives ratio, limited identification of premalignant and malignant disease of cervix.

On the contrary, study of the smear with phase contrast microscopy allows simultaneous cytohormonal evaluation, indication of components in vaginal eco-system, determination of menstruation cycles and oncological diagnosis.

The method allows the assessment of unstained and fixed cell samples immediately after the sample was taken, which is very important in everyday practice. Such a short diagnosis time is very much appreciated by patients who are spared the stress which is usually connected with the examination. It has also a therapeutic meaning, especially in cases of serious illness. Moreover, observed cell samples have to be neither stained nor dyed which yields, on one hand, more reliable and accurate diagnosis, and on the other hand, fast degradation of the sample, which dries up after 3-4 hours. The use of digital methods of recording microscopic images provides a solution to this problem. Because the evaluation of hundreds or thousand of images is a tedious task, an automated image analysis is indispensable.

Our intention was to present a method that will simplify evaluation of images obtained from the phase contrast microscope, tthat automatically analyzes images generated during an examination, and that presents the physician with only regions of interests that contain objects essential for oncological screening, namely epithelial cells. Non-epithelial elements such as granulocytes, semen, lymphocytes, erythrocytes, vaginal microflora (bacterias or viruses) and artifacts that result from false sample preparation obscure the diagnosis and, as irrelevant, should be removed from images.

Results

A statistical algorithm was proposed to characterize and recognize patterns in epithelial cells. It follows a three stage process. At first, a set of features are extracted. Then, the most discriminant features are selected and used to classify a given image. Finally, post processing steps refine the classification results and an active contour model is employed to locate membranes of epithelial cells.

The classification performance was tested with three different classifiers : Fisher Linear Discriminant (FLD), Kernel Fisher Discriminant (KFD) and k-nearest neighbor classifier. The correct classification was equal to 84.8%, 86.8% and 84.6%, respectively and showed that KFD provided the best accuracy. An additional algorithm applied edge detection, ridge following, contour grouping and FLD to detect abnormal cells.

Evaluation of the algorithm's performance and comparison with alternative approaches show that the method is reliable. By presenting only images or their parts that are diagnostically important, the method unburdens the physician from massive and messy data, indicates abnormal structures and, in that sense, supports diagnosis of cervical cancer. Therefore, it can be used to identify epithelial cells as well as nuclei on phase contrast images.

We are open to inquiries from vendors, producers of microscopes and companies interested in this application.

Image 1. Detected cells are marked in green. An additional algorithm detects nuclei and marks them in red. The combination of both approaches shows the physician regions of interests (cf. third row). Both methods can be also applied separately (cf. first and second row).