Machine Learning

Figure 1. The image of droplet condensation and the corresponding image treated by image thresholding (Scale bar is 100μm).


I need an efficient and accurate method to annotate the droplets contained in the Time-lapse images. It is challenging for the "traditional image analysis" methods, such as image thresholding (Figure 1). Because of few things:

  1. Various focus between droplets
  2. Various intensity between regions
  3. Segmentation difficulties
Figure 2. The schematic of the machine learning of droplet annotation (Scale bar is 100μm).

To overcome the challenge, I adopted the machine learning approach. I found a open-source model named Stardist, which is aimed to detect the Star-shaped polygon1. I only need to prepared the training sets to get the model (Figure 2).


Because my image is composed of the droplet at various size, the majority is ~10-6 m and there are a few droplet at 10-4 m. In consequence, the training volume of the "big droplet" was less than the "small droplet", led to the inaccurate model if I did not enhance the training volume of "big droplet". However, it will take me tremendous effort to prepared the enough training volume for the "big droplet".


Here I would like to share a little trick. First, I prepared a few training set and trained the model. Second, I used the trained-model to annotated the "small droplet". Therefore, I only need to annotate the "big droplet" to prepare the good enough training sets.

Figure 3. The image of droplet condensation and the corresponding annotation of the image by machine learning method (Scale bar is 100μm).

The result is satisfying (Figure 3)!


  1. Schmidt, U., Weigert, M., Broaddus, C., & Myers, G. (2018). Cell detection with star-convex polygons. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11 (pp. 265-273). Springer International Publishing.