AI-driven analysis for real-time detection of unstained microscopic cell culture images

Abstract

AI-based image recognition has significantly advanced the analysis of tissues and individual cells both in the context of translational studies and diagnostics. To date, recognition is primarily based on the identification of certain cell characteristics (e.g. by staining). The morphological assessment of unstained cells holds additional potential, as it allows for virtually real-time assessment without the need to manipulate the cells. This facilitates longitudinal observations, as required for drug testing, and forms a basis for autonomous experimental execution. A semi-automated cell culture system (AICE3, LabMaite) was used to culture myeloid leukemic cell lines (K562, HL-60, Kasumi-1). K562 cells were treated with hemin and PMA to induce erythroid and megakaryocytic differentiation, respectively. Cell images were acquired using automated bright field microscopy. Images were used to train an AI model using an NVIDIA DGX A100 GPU with Ultralytics YOLOv8. Morphologic features were extracted using RedTell. The model reliably distinguished K562 cells from HL-60 and Kasumi-1 using >400 images per class (average >15 cells/image). Bounding boxes were generated correctly (mAP@.5 >98%); precision and sensitivity exceeded 97%. Validation on an external K562 dataset confirmed these results. Classification of all three cell lines achieved >97% sensitivity/specificity and 94.6% precision. To test drug response, we used YOLOv8-s to distinguish untreated K562 cells from those undergoing erythroid or megakaryocytic differentiation (n >3,000 annotations). Precision, sensitivity, and specificity were >95%. RedTell identified 3 of 74 morphological traits contributing significantly to class separation. We demonstrate accurate, near real-time detection of unstained cells, enabling future AI-based drug testing.