(PA-305) Computer-assisted AI to Detect Myeloma Cells

Multiple myeloma remains an incurable disease. If we advance plasma cell/myeloma cell detection by novel approaches, we may be able to work towards a cure. Our team set out to develop computer-assisted AI to identify individual myeloma cells on modified Giemsa-stained slides. Immunophenotyping (CD138/CD56) confirmed the nature of the cells as myeloma cells.


Methods:   

Overview: The computer was trained with 100% myeloma and 100% normal blood cells, followed by different dilutions of myeloma cells and blood cells. Finally, spiking experiments were carried out with up to 1 myeloma cell per 1 million normal cells.  The recognition of myeloma cells was then confirmed in myeloma patient samples.

The detection method is based on a combination of advanced optical microscopy measurement of the stained slides combined with a computer program that identifies the cell types and performs the statistics.

Optical system: Digital imaging evolved during the last few years. Currently, it is mainly based on what is called whole slide imaging (WSI), which is basically a system that can scan the whole slide with a given magnification in color, store the data and allow an expert to overview the results on the computer screen.

For our method we can use a conventional WSI system, but we also use a new type of a microscopy-based imaging system that allows to measure the full visible light spectrum at every pixel of the object. This spectrum contains much more information relative to the color, and acts as a fingerprint of the underlying tissue. This information must be processed and analyzed in order to extract the relevant information.

Computer program:  The computer program that analyzes the measured information contains few modules. The first one use AI algorithms that we adopted for identifying the nuclei of all the cells in the image. We found that the nuclei, for this application, provide significant information. Then, different features of the nuclei are calculated such as the size of each nucleus, the shape of it, circularity, roughness of the nucleus envelope, heterogeneity of the chromatin in the nucleus and more. Some of these calculations are performed using conventional image-processing tools, and others use AI.

 

Results:   

We tested different slides, including a mixture of normal and myeloma cells and patients’  bone marrow. The results confirm the validity of the methods and provides good results in identifying myeloma cells.

Finally, spiking experiments were carried out with up to 1 myeloma cell per 1 million normal cells.  The recognition of myeloma cells was then confirmed in patient samples with a ratio of myeloma cells to normal cells of 1:100,000 and even 1:1,000,000. The analysis confirms the ability of the system to measure even a small ratio as 1 myeloma cell to 10⁶ normal cells.

 

Conclusions:   

Computer-assisted AI is able to identify myeloma cells in blood and bone marrow. The method is ready for further application in patient samples.