Artificial Intelligence (AI) has an enormous computational power to process large, complex quantities of data. And yet, to harness this power, seasoned and astute computer scientists must leverage both their knowledge and experience to apply the right techniques for the best results every time.
Current limits to image analysis techniques for cancer staging
At Definiens, Principle Research Scientist, Dr. Nicolas Brieu, was faced with the challenge of developing image analysis techniques to segment and quantity Immunofluorescence (IF) images from MIBC patients.
The TNM staging system is a widely established method of staging malignant tumors in different indications. But in some cases, such as the highly-aggressive muscle-invasive bladder cancer (MIBC), the separation can be further improved by incorporating additional parameters.
Why use machine learning & deep learning to classify tumor stages?
Advanced AI approaches such as machine learning and deep learning can be applied to improve TNM staging. Choosing how and when to apply them, however, depends on the biological premises. The premise of the Immuno-oncology (IO) approach studies the interactions between the tumor and selected immune cell populations.
The IO - Immuno-oncology method detects cells of different cell populations based on their biomarker expressions and additional robust tumor classifications to localize different cells in relation to the tumor itself. As a result, the densities of various cell types can be measured in different regions.
On the other hand, the 'Tumor budding' approach also requires the tumor classification, but is then followed by a morphological analysis of small cancer clusters, or 'tumor buds.'
Developing a common and robust accurate image analysis solution to classify the tumor cells is a critical first step to ensure accuracy of subsequent analysis. So, Dr. Brieu and Dr. Peter Caie (St Andrews University) decided to combine their years of experience to find the best technique combinations. Caie used a combination of techniques using the Random Forest learning classification method in addition to Convolutional Neural Networks (CNN), which are inspired by biological processes and used to analyze visual imagery. For cell detection and to overcome the variability in size and texture, Dr. Brieu focused on detecting the nuclei. In this case, he used the well known Proximity Map, while leveraging the local Surface Map to secure robust and accurate results.
Want to know what their results were? Learn more about deep learning and tumor budding in our webinar with Dr. Brieu and Dr. Caie to learn the techniques they used to improve TNM staging and the improved prognostic stratification they observed.