https://doi.org/10.1016/j.optlastec.2024.111992

Hong T., Peng M., Kim Y., Schellhorn H. E., Fang Q. (2025) Optics & Laser Technology, Vol 181, Pg 111992, Issn 0030-3992, https://doi.org/10.1016/j.optlastec.2024.111992

Image-based cell profiling offers high-quality profiles (i.e., morphological phenotype) after high-throughput microscopic image acquisition. It requires large-scale image analysis and processing ability, and deep learning has recently demonstrated outstanding performance in this field. However, deep learning heavily relies on conventional manual annotation, which becomes a critical bottleneck for the entire profiling workflow. This study develops an annotation-efficient self-supervised active learning pipeline for images acquired by high throughput imaging flow cytometry. First, a fitting pretext task using optical flow and variational autoencoder algorithms is proposed to pre-train the model. The supervisory is derived from the motion cues of cells with zero manual annotation. Then, an active learning cycle selects a small set of samples to annotate, which iteratively achieves an optimal performance with fewer annotated samples. The sample selection criteria consider informativeness and representativeness from static and spatiotemporal features. Once one cell is selected to annotate, it is removed from the unannotated data pool, further reducing human effort. The pipeline is used for a lensless optofluidic imaging flow cytometry and experimentally evaluated by testing three components of biological samples in urinalysis (erythrocytes, leukocytes, and budding yeasts). It shows comparable performance but with only an average of 30–40% annotation workload compared with fully supervised training. The generated image-based profiles can be used for downstream analysis. These results indicate that the proposed training pipeline yields high performance and efficiently decreases the annotation burden for implementing deep learning in image-based cell profiling.

Keywords: Annotation-efficient, Self-supervised, Active learning, Optical flow, Spatiotemporal feature

https://doi.org/10.1016/j.optlastec.2024.111992