Medical image segmentation development process

Medical segmentation: unlocking valuable information

Medical image segmentation plays an important role in medical image analysis for clinical research and practice purposes. It enables more effective diagnosing and planning of treatment of lesions of oncological origin, as well as those affecting the cardiovascular system, e.g. coronary artery disease.

Image segmentation, which is the process of dividing an image into specific areas, allows for in-depth analysis. Thanks to segmentation you may generate a selected anatomical structure in three dimensions and separate it from the adjacent tissues. Despite the progress that has been made in this field over the past 20 years, many challenges remain. These challenges mainly concern obtaining precise and accurate results, but not only. Another hurdle is building smart algorithms that work on both healthy and diseased organs and tissues.

At Graylight Imaging we have over 15 years of experience in developing systems for medical imaging. We have created a methodology and process that enables us to achieve high-quality results.

How to perform with quality and precision?

Medical segmentation is only a seemingly easy process. Several components affect the final effect. Of course, there are factors beyond our control or the control of our clients, just to mention the availability of large, diverse datasets. However, there are challenges involved in medical image segmentation we can manage. Below you might find some of the improvements we have implemented based on lessons learned from past projects.

Effective segmentation methodology by Graylight Imaging

Over the past years, our specialists have been ensuring that the models and algorithms we develop perform with high quality and precision.

• Cooperation of specialists from various fields. Our interdisciplinary teams consist of biomedical engineers, electro radiologists and physicists. Of course, they are supported by specialists in machine learning and software development. They work together on the development of segmentation models.
• A team of medical imaging specialists. The team consists of people with competences that allow them to understand both medical and technical issues. This highly specialized group of experts works on highly accurate 3D models of anatomical structures, constantly exploring the secrets of human anatomy. This is where the difficult segmentation process begins, which is essential to build ML models.
• We work on diverse data. The issue of biased training data is widely acknowledged. However, it’s not just about diversifying the data in terms of patient race or gender. It’s crucial to ensure a high level of generalization for the algorithm’s performance. That’s why we work with data from different modalities and originating from devices made by different manufacturers.
• Anatomical diversity. Additionally, we have perfected the imaging precision on anatomically diverse organs. Coronary artery segmentation is a great example here, as these structures are featured by high anatomical variability.
• Cross-consultation. This is an important element of the process that brings measurable benefits. Members of the team working on segmentations review each other’s results. This approach makes it easier to find minor errors made by a co-worker. But there is more to it. One of the benefits is also the opportunity to learn from each other.
• Our own segmentation software. Finally, at Graylight Imaging we have developed an original solution that meets our high requirements. It is adapted both to our needs and the requirements for medical devices.

Medical segmentation. Increase the difficulty level

Years of work on medical imaging systems have brought tangible results. We started with high-quality examinations, examinations without anatomical anomalies. In the next step, as time went on and we gathered more and more experience, we increased the difficulty. Today, projects involving the development of segmentation models are handled by teams of highly specialized professionals combining various competences.

The real measure of success

Our experience is reflected by successes in both commercial and R&D fields. The models we developed automatically segment the coronary artery tree while maintaining a reconstruction accuracy of over 90% (Sørensen-Dice coefficient of about 0.9) [1]. Another one automatically segments the brain tumor along with its subregions (edema, tumor enhancement and necrosis). It also provides data on volume measurement and RANO (Response Assessment in Neuro-Oncology) [2]. These are just a few examples of our capabilities.

References:

[1] For more information go here: https://graylight-imaging.com/cardiovascular-imaging/ 

[2] Nalepa J. et al., Deep learning automates bidimensional and volumetric tumor burden measurement from MRI in pre- and post-operative glioblastoma patients, Computers in Biology and Medicine, Volume 154, 2023, https://doi.org/10.1016/j.compbiomed.2023.106603