Aleksandra Walczybok, a graduate of our Faculty, won this year’s edition of the CreatiWITy competition in the engineering thesis category. Her victory came thanks to an original application supporting the analysis of cytological specimens, designed to assist pathomorphologists. Congratulations!

Data engineering in medicine

During her first-cycle studies in Systems Engineering, the graduate focused on data engineering, working with models and machine learning. This interest also shaped the topic of her engineering thesis. Medical applications naturally became the direction where data analysis can genuinely support specialists.

Her engineering thesis, titled “Design and Implementation of a Machine-Learning-Based System Supporting the Analysis of Cytological Preparations”, concerns an application analysing microscopic images of cells collected during cytological examinations, which play an important role in the prevention and early detection of cancerous changes and the HPV virus.

Processing real data

The thesis was supervised by Dr. Łukasz Jeleń, whose research interests align with the subject of the project. Thanks to this collaboration, the student gained access to real anonymised data – collections of cytological specimen images collected and shared by the Department of Pathomorphology at the Pomeranian Medical University in Szczecin. As she emphasises, this was one of the greatest advantages of the project – working with demanding real-world data rather than publicly available datasets from the internet.

– The aim of the application is to support and improve the work of pathomorphologists. The process often takes a long time, especially in smaller centres where there are fewer laboratories. The application is designed to speed up this process – explains Aleksandra Walczybok.

The system developed by the author introduces three categories: healthy, suspected HPV infection, and high suspicion of cancer.

Multi-stage image analysis

The solution is based on multi-stage analysis of the specimen image uploaded to the application. The process begins with cell detection. – I crop a specific cell fragment from the image and then it passes through two separate paths – explains Aleksandra Walczybok.

The first path uses convolutional neural networks, primarily applied to image analysis. The model classifies the cell directly based on visual features. – It works similarly to classic image recognition examples: identifying whether there is a dog or a cat in a photo – here the mechanism is analogous – she adds.

The second path begins with segmentation, meaning identifying the area of the cell and its nucleus. Based on this, the system generates masks that enable the calculation of morphometric features, such as the nucleus-to-cytoplasm ratio or the cell perimeter. These are parameters that cytologists pay close attention to in their work. The extracted morphometric features then become input for classical machine learning models such as Random Forest, which predicts the cell class.

– I wanted to take advantage of the complementarity of models. Each performs best in slightly different cases and makes different errors. That’s why I combined both paths to increase accuracy. Both models generate probability vectors, which I then combine into a final assessment for each of the three categories – explains the winner.

This approach not only categorises cells but also provides pathomorphologists with additional information about their parameters. – The application displays defined morphometric features, so a specialist can quickly see, for example, the size of the nucleus or the nucleus-to-cytoplasm ratio.

The project was created as a tool supporting specialists. – The specialist must make the final decision about the cell class. Such a tool is meant to assist – she explains.

An important element of the thesis is also the use of AI explainability methods. In the case of convolutional networks, the application highlights the areas of the image that had the greatest impact on the model’s decision. If the system focuses on the cell nucleus, the result is considered particularly valuable from a diagnostic perspective. In the second path, the LIME approach is used to present a diagram showing the importance of morphometric features. These mechanisms allow experts to assess whether a prediction is justified or may result from random correlations.

– We cannot trust a model 100 percent. That is why explainability is so important – it allows us to understand the basis on which the system made its decision – adds the WIT graduate.

The system operates not only at the level of individual cells but also predicts the category for the entire slide. The assessment of the specimen is based on aggregated results and probabilities assigned to cells, which helps better handle ambiguous cases.

The application also allows data archiving and sharing slides with other pathomorphologists. – We can return to previous examinations of a patient and check what has changed over time. It also supports consultations and the creation of a knowledge base.

The project is currently not intended for general use. As the author – who plans to continue her studies at WIT in the Artificial Intelligence programme – explains, she is currently planning to prepare a scientific article on the subject.