Prof. Halina Kwaśnicka: Artificial intelligence in Polish medicine? Sure, but give us access to data

The Department of Computational Intelligence, Faculty of Computer Science and Management, Wrocław University of Science and Technology, where I work, collaborates with physicians to create for them useful systems.

I led three scientific projects carried out jointly with doctors from Wrocław Medical University: “Construction of a comprehensive system for computerized analysis and measurement of excess cerebrospinal fluid in the midbrain to support radiological and neurosurgical decisions” (2006-2008); “Development of a comprehensive system for automated diagnostics of small vessels based on computerized analysis of capillaroscopic images” (2007-2009); “Application of methods of computerized analysis of images for the assessment of the expression of HER-2 receptor in breast cancer cells” (2011-2014).

The scope of our projects is presented below.

Project 1: Dementias and strokes

Cerebral atrophy is one of the features of dementias and other neurodegenerative conditions, including physiological brain aging. We were interested in determining the relationship between measurements of cerebral atrophy and the level of dementia based on neuropsychological testing. We conducted neuropsychological testing and CT exams [Editor’s note: computed tomography] in a group of 68 patients diagnosed with cognitive dysfunctions caused by dementia.

No statistical correlation between the results of neuropsychological tests and measurements of cerebral atrophy (both cortical and subcortical) was found. Measurements of generalized cortical and subcortical atrophy can be used in therapeutic process to compare the progress of atrophic changes with changes in the results of neuropsychological tests.

We have developed a medical support system, including segmentation, extraction of features and decision-making processes. Various techniques are used in the system: fuzzy logic, shape descriptors, classifiers and structure for automated description of images. We have proposed a simple method for early detection of cerebral stroke based on CT images.

In all of these research efforts, especially concerning early stages of stroke, we held a small sample of data – from 40 to 70 images, derived from a single tomograph and a single team of physicians. It is definitely too little to determine whether the methods or systems that we have developed are suitable for general use.

Project 2: Capillaries

Capillaroscopy is a branch of medicine focused on analysis of changes in the capillaries. There are many important aspects of the capillary image analysis: thickness, shape, distribution and density of the capillaries [Editor’s note: capillary – a very thin tube, here: a thin blood vessel]. The analysis of such images is helpful in diagnosing many rheumatoid diseases.

Due to the fact that at that time we were working on methods for automatic annotation of images based on their visual content, we applied this approach to actual capillaroscopy data. It was interesting, because it did not require identifying specific features of the capillaries seen on the image.

In research based on the available, small data sample, we achieved the accuracy of ca. 77 percent It is a good result. Further research was to focus on acquiring a significantly larger set of capillary data and adding new potential diagnoses to the dictionary. However, we could not verify this approach, because we did not have real-life data.

The problem seems to be the lack of time and young people to work for the money offered

We also proposed a method to measure the curvature of the capillaries which worked with 84-percent accuracy on real-life data. But again, the problem with the size and quality of the data set recurred: the old capillaroscope that we had was unable to take good, professional pictures of the capillaries. The biggest problem was the identification of branching and superimposed vessels.

One of the most important indicators in capillaroscopy is the thickness of the capillaries. Based on this, the capillaries can be grouped into three categories: healthy capillaries, those with enlarged loops, and megacapillaries [Editor’s note: vessels with the diameter much larger than the diameter of healthy capillaries]. That is why we focused on automation of the analysis of capillary thickness. We experimented with various classifiers. The highest accuracy reached 97 percent, which can be considered satisfactory – a system working with such accuracy could support diagnosis.

We also focused on automatic analysis of the distribution and density of the capillaries. A new approach to the detection of the ischemic areas that we have proposed uses the analysis of histograms and classification.

We have also proposed an innovative, semi-automated method for capillary tracking. Methods for automated analysis (with various accuracy levels) have been developed for all capillary features that are important in image-based diagnosis. The problem was to develop a universal segmentation method to deal with all anomalous capillaroscopic images.

Project 3: Breast cancer

Human epidermal growth factor receptor 2 (HER-2/neu) is an established biomarker for diagnostic and prognostic assessments of breast cancer. Patients are diagnosed with breast cancer by accurate identification of cell membranes, visualized as “overexpressed HER2/neu”, in the images obtained from adequate histopathological preparations. A traditional “manual” analysis of those images is a tedious and time-consuming exercise, and automated analysis is a difficult process. It is necessary not only to identify overexpressed cells, but also to determine how numerous they are as compared to all cells in the image.

To segment these structures, we have proposed appropriate approximations of a fuzzy set. This research has been continued as analysis of HER-2/neu membrane cell junctions by analyzing the proposed coefficients of shape.

The research underlies the introduction of a complete procedure to diagnose breast cancer. We completed the project with a proposal for a complete system for automated support of the diagnosis of breast cancer, paying special attention to accurate identification of the boundary between the second and the third levels of this disease. It is very important for therapy, and expert diagnosis of cases that are very close to this boundary is very difficult.

Since we held a small amount of data for system training and testing (they came from a single medical facility), its complete and reliable verification was not possible.

Little time, little money

We have continued joint research with physicians as part of the research conducted by the workers of the Department of Computational Intelligence, albeit not on a large scale. Before summer holidays in 2018, we held several meetings with teams from Wroclaw Medical University. We can see the potential of artificial intelligence in medical tasks, and the physicians also understand the need for and sense of such collaboration. The problem seems to be the lack of time and young people to work for the money offered.

Informal collaboration with physicians addressed several problems.

The team led by Martin Tabakow, PhD from “AKSON” Rehabilitation Center for Treatment of Spinal Cord Injuries in Wroclaw was searching for a model based on fuzzy control that could be applied in a bionic hand. We focused on its real-time operation (clenching fists).

Based on real-world sEMG (electrical activity produced by muscles) signals collected from patients with an amputated hand, we proposed a method for the operation of a potential controller based on sEMG signal data and data from the installed pressure sensor. A controller model has been created which is compatible with the assumptions for the bionic hand.

The collaboration with AKSON allowed us to track even small changes in patient’s condition, which is important especially during rehabilitation. We have used data from the sEMG signal during the exercises performed by patients with spinal injuries as well as the physical therapists’ assessment of the “quality” of how the exercise was performed. We have proposed a new, objective factor for the assessment of the condition of rehabilitated patients, called the “neuromotor factor”.

We returned to the problem of breast cancer diagnosis in 2018. This time, we analyzed the segmentation and identification of cancer cells from Ki-67 testing [Editor’s note: Ki-67 is a protein which, when detected in cells, indicates that such cells divide, which may suggest cancer.]. An innovative method for counting cells that are glued together or superimposed on each other was proposed; such cells are “counted” by a trained convolutional network [Editor’s note: a type of neural network]. The results obtained for the data tested are better than the results of commercial software.

Professor Jerzy Sas was the leader of research which, although not directly related to medical diagnosis, has proven useful in Polish healthcare facilities. The research focused on algorithms for the identification of hand-written medical texts, which is important for digitization of medical records.

An equally important problem is to develop a speech recognition system in the medical setting. Such a system should make it possible for, e.g., a radiologist to orally describe the image analyzed. The system would save the “description” as text file.

The potential is huge, but…

The potential of artificial intelligence in medicine is huge, but depends on many factors. A fundamental element hindering or even preventing the emergence of good solutions is the lack of large and diverse data sets that could be made available to researchers and interested companies.

Another problem includes the lack of educated young people willing to work with research projects. The salaries of assistants at Polish universities are several times lower than the salaries of IT specialists in companies. Companies often even poach employees, including those from universities. The cooperation between industry and science is also insufficient.

The text is an adaptation of the paper by professor Halina Kwaśnicka entitled “Comments on the strategic artificial intelligence program in Poland. Development of machine learning methods with reference to medical applications and threats arising from the pace of development of artificial intelligence around the world” (Wrocław 2018), prepared for the National Information Processing Institute.

Skip to content