Abstract:

Malaria remains to be a global health challenge causing half a million deaths annually. With imaging processing and deep learning, many approaches have been proposed to analyze Microscopic Blood Films automatedly; however, the presence of artifacts increases the difficulties of digital pathology. This project mainly focuses on stain artifacts in Giemsa stained thin blood smear. We aim to verify the feasibility of image processing method and machine learning models that could help remove artifacts in appointed images. The critical challenges of this project are lacking cell-level labels, corresponded supervised dataset, and the difficulty of separating parasites and artifacts which looks similar.

The proposed image processing algorithm is concerning both color difference and area difference of artifacts and cells. The algorithm divides the artifacts into two sets (inside and outside the cells) and removes them under similar ideas by selecting out all artifacts, making masks for the selected area, and then inpaint the mask. However, due to the random feature of artifacts, it is impossible to find a suitable threshold for all types of artifacts, and there are too many of them, so we conclude this algorithm to be inefficient. We have applied the CycleGAN framework to our dataset and fine-tune it to be suitable for our problem, and we find this method is effective in removing small artifacts but may perform weaker on larger ones.