Abstract
The quantification of biological assays, such as plaque and microbial assays is essential in virology and microbiology research. However, low-contrast images of stain-free samples are difficult to segment accurately and manual labeling is time-consuming. To address these problems, we present a weakly supervised framework for automated biological assay assessment. First, we collected and constructed weakly supervised datasets for viral plaque and microbial colony segmentation using point and bounding box annotations respectively. Then, we proposed an adaptive region-growing algorithm that generates mask annotations, reducing annotation burden. We adapted and fine-tuned automatic Segment Anything Model (SAM) to for biological specimen segmentation, demonstrating improved accuracy across diverse assay types. Moreover, we also validated our method on live cell segmentation. Finally, we applied our model in antiviral compound assessment and achieved comparable results to manual assessment. In summary, our framework provides an efficient and automated solution for biological assay quantification, reducing annotation burden while maintaining accuracy.