Session: 803. Emerging Tools, Techniques, and Artificial Intelligence in Hematology: Poster II
Hematology Disease Topics & Pathways:
Artificial intelligence (AI), Emerging technologies, Technology and Procedures, Measurable Residual Disease , Machine learning
Current CPC enumeration and isolation methods face drawbacks like low cell recovery rates, poor purity, and reduced viability. Physical methods lack specificity, and filtration can deform or miss smaller CPCs. Immunomagnetic assays, like CellSearchâ„¢, can damage cells and require extensive enrichment processes and manual handling, making the process cumbersome and error-prone. These challenges hinder clinical and laboratory applications.
We address these limitations with an enrichment-free, gentle, and highly sensitive method for detecting and isolating CPCs. Our core innovation is a transformer-based deep learning model that combines CD138+ surface expression and morphological images of cells flowing through a novel, highly parallel microfluidic setup, distinguishing CPCs from blood cells without explicit negative selection. While healthy CD138+ CPCs likely exist in circulation at very low concentrations (~5 CPC/mL), CPC counts in MM patients are significantly higher, making CD138+ CPC count a reliable measure of the burden dynamics.
A transformer-based deep-learning model was trained on 13,308 U266B1 cells and roughly 3 million Peripheral Blood Mononuclear Cells (PBMCs) for negative control. The model operates on a rolling fixed size context window that combines 24 consecutive expression and morphology images captured at spatially separate instances to compute the most likely trajectory of target cells. A separate deep-learning network is trained to estimate a tunable sensitivity threshold (T) to balance detection and false-positive rates.
Preliminary validation results on the core algorithm are reported. Three independent runs were performed with U266B1 cell line MM B lymphocytes. High-speed video recording of the MM cells was used to partition the runs into segments with exactly 100 MM cells. Twelve segments were analyzed for computing the detection rate. Two separate negative control runs were performed, each with roughly 3 million PBMCs isolated from 1 mL of whole blood from healthy donors. The average run-time was 4.5 hours with no manual intervention.
Adjusting the sensitivity threshold T to allow for at most 1 CPC in 1 mL of healthy blood (T1), the detection rate of MM cells in the positive samples was 41.6% (95% CI ± 3.8%). With the threshold adjusted to allow for 2 CPCs in 1 mL of healthy blood (T2), the detection rate increased to 81.2% (95% CI ± 5.6%). The CPC count in healthy blood is well within the reported range for healthy individuals.
For a more realistic scenario, we spiked roughly 1000 MM cells into 100K PBMCs suspended in 10ul of FACS buffer. Five separate segments with exactly 100 MM cells in each segment were recorded, as confirmed visually by analyzing a high-speed video recording. Using detection threshold T2, the detection rate was 78.8% (95% CI ± 7.8%), in line with the MM-only run above.
These preliminary results demonstrate the ability of transformer-based AI algorithms to efficiently combine sequences of images of different modalities to improve the detection performance in detecting extremely rare cells at scale. As our first application, we demonstrated its performance for CPC enumeration and isolation in MM. To develop a robust assay, ongoing clinical experiments aim to quantify cell loss during sample preparation and validate performance with patient samples at different stages of MM. Beyond CPCs, our models show promise in detecting and isolating other rare cell types, including circulating solid tumor cells and stem cells from breast milk and amniotic fluid, establishing them as valuable tools in rare-cell biology.
Disclosures: No relevant conflicts of interest to declare.