B2 - Simulation-Guided Adaptation of SAM2 for Generalizable Calcium Activity Segmentation in Non-Excitable Tissue Imaging

Introduction: : Calcium signaling regulates key biological processes such as regeneration, migration, and differentiation, particularly in non-excitable cells where Ca²⁺ transients influence critical decisions about cell behavior and fate. Accurately identifying Ca²⁺-activated cells in tissues and organoids is essential for linking single-cell signaling dynamics to large-scale biological outcomes. Yet, analyzing time-lapse imaging data is hindered by dense packing, low contrast, and faint signals in wide-field microscopy. Further, few tools are easily generalizable to new applications. Here, we introduce a new simulation-driven optimization strategy that leverages the performance of powerful foundational models like Segment Anything Model 2 (SAM2) to accurately quantify and analyze Ca²⁺-active epithelial cells, with new methods to bridge the synthetic-to-real data gap.

Materials and

Methods: : We generated 50 synthetic videos each comprising 180 frames using a biophysical Ca²⁺ model that simulates a Drosophila wing disc that is imaged with simulated fluorescence microscopy, providing ground-truth masks and incorporating realistic artifacts, including background autofluorescence, spontaneous luminescence, Poisson/Gaussian sensor noise, non-uniform illumination, partial defocus, and edge softening.
After applying random point prompting on masks with 5x5 erosion to reduce edge bias, we divided the synthetic videos into an 80/20 train/test split. We then fine-tuned the mask decoder of SAM2 ("sam2.1-hiera-small" with frozen encoders), optimizing for Intersection over Union (IoU). Performance was evaluated on held-out simulations and 200 real experimental frames using mean IoU (mIoU), Dice coefficient , and F1 score, which provide complementary metrics of segmentation accuracy and offer a balanced assessment of false positives and false negatives.

Results, Conclusions, and Discussions:: The fine-tuned SAM2 demonstrated excellent performance on experimental data, achieving an mIoU of 0.8, a Dice coefficient of 0.89, and an F1 score of 0.89. The high mIoU and F1 score indicate that our model reliably detects calcium-activated regions and rarely misclassifies autofluorescent areas, despite substantial imaging artifacts. Notably, the model's performance matched the previous SAM2-path, which was specifically developed for pathology segmentation tasks. This new approach provides three main advantages: (1) Robust single cell segmentation in complex calcium imaging conditions, (2) Reducing hallucination in low-signal regions, (3) Effective transfer learning from synthetic real data.

Acknowledgements and/or References (Optional):: Current results pinpoint specific discrepancies—z-stack blur, star-convex cell shapes, and heterogeneous background fluorescence—that limit accurate automatic analysis. Simulation-driven fine-tuning can quickly endow a foundation segmentation model with strong performance in a new domain. As next steps, several extensions are being incorporated to resolve several challenges toward robust deployment on experimental calcium imaging data.
Ultimately, a hybrid pipeline that combines physically grounded simulation with minimal real-world supervision shows promise in unlocking large-scale, unbiased quantification of intercellular calcium signaling dynamics in complex non-excitable tissues.