LSTM-based Effector-Target Cell Classification in Ex Vivo Co-cultures Treated with Immunotherapies

This work was done by Dr. Qibing Jiang at University of Central Florida under the mentorship of Dr. Wei Zhang and me.

 

The development of TLCellClassifier is motivated by the growing need for precise, automated analysis tools to support high-throughput, image-based preclinical assays in hematological malignancies like multiple myeloma. As immunotherapies gain traction in clinical treatment, determining optimal choice of therapy for individual patients remains a significant challenge. Ex vivo assays that mimic the immune-tumor microenvironment using patient-derived tumor (target cells) and immune cells (effector cells) offer a powerful platform to screen therapeutic responses dynamically. However, these assays generate massive time-lapse brightfield image datasets capturing complex interactions between multiple cell types over several days. Accurately detecting, tracking, and classifying thousands of tumor and immune cells in each frame requires a specialized approach. TLCellClassifier tackles this challenge by combining convolutional neural networks (CNNs) for robust cell detection, metric learning for reliable cell tracking, and a two-stage LSTM framework for accurate cell type classification. This integrated approach enables scalable, high-fidelity analysis of live-cell behavior in immunotherapy assays, advancing the goal of personalized treatment selection in cancers like multiple myeloma.