Fluorescence microscopy coupled with acridine orange/propidium iodide (AO/PI) dual staining is widely used to distinguish viable, apoptotic, and necrotic cells, yet its interpretation relies on manual inspection, which is time-consuming, subjective, and dependent on expert judgment. Conventional machine learning and convolutional neural networks require extensive annotated datasets and are often task-specific. Multimodal large language models (MLLMs), which interpret visual and textual data within one cohesive framework and generate structured outputs aligned with domain-specific criteria, offer a new paradigm for biomedical image analysis.

Recently, a team from Atilim University at Türkiye, published a research article, “Artificial intelligence‐enabled cytopathological interpretation of fluorescence microscopy images via a multimodal large language model” in Quantitative Biology. The team evaluated how well an MLLM, Claude 3.7 Sonnet, interprets AO/PI-stained fluorescence images of doxorubicin-treated MCF-7 cells. A total of 500 images from one control and four dose groups were analyzed. Ground truth was established by three experts using a standardized 15-parameter cytopathological evaluation checklist, and model outputs were compared against expert judgments across cell counts, viability, and apoptotic and necrotic indices. The entire dataset was processed within 2 hours, a substantial gain in analytical throughput over manual evaluation.

Figure 1 compares expert and MLLM assessments of six key cytopathological parameters. (A) Total cell counts agreed closely across all dose groups, and (F) necrotic indices showed a consistent dose-dependent increase. However, (B) the model systematically overestimated viability, and (C, D, E) it failed to reproduce the expert-observed peaks in early, late, and total apoptotic indices at intermediate and high doses. Full convergence with expert scores was reached in 51.1% of parameter-dose combinations, and intra-model repeatability reached 92.5%. The model achieved strong discrimination for necrotic cells (AUC = 0.84) but near-random performance for early and late apoptosis (AUC = 0.52 to 0.55), reflecting limited spatial and contextual inference.
DOI
10.1002/qub2.70042