In HCS Pharma we are specialized in 3D cell culture and we have developed high skills in 3D imaging thanks to the help of our partner Molecular Devices. We are used to using powerful ImageXpress Micro Confocal Systems to investigated 3D cells morphology in HCS context from fluorescence pictures. The very first step to collect morphological data is a precise segmentation of objects in the 3D picture. There are many challenges to do that! First, there are different resolutions in the picture, one for X and Y axes (picture plan) and another for the Z axes (the height). It’s an important point for a good picture 3D reconstruction. Secondly, as BIOMIMESYS® is a translucent medium, we see several kinds of “blur” in the picture. To be short, it complicates the selection of all pixels belonging to a single object.

To overcome all these issues and to accelerate picture analyzing, we have launched the VisuAI project with several partners. First results has been publish in “An evaluation of computational learning-based methods for the segmentation of nuclei in cervical cancer cells from microscopic images” by Maylaa et al.

This article presents the performance of three machine learning architectures for 3D pictures segmentation. We find very good segmentation for 2D pictures, event when we observe complex mixing of different objects.

Example of segmentation : (C) original picture in DAPI (D) Nuclei segmentation with machine learning

Another interesting result was the very low rate of false positive detection. It shows that our algorithms are robust against noise and cellular debris.

Visual representation of the nuclei segmentation results on the original image using two ML methods: True Positive in red, False Positive in green and False Negative in blue.

We also shown that the generalization of 2D segmentation to 3D segmentation is not obvious. It needs manual labeling of many pictures at different Z, which is time consuming. It will be interesting for experiments which use same kind of pictures. In R&D context, it’s not the case. It’s why we are working now on some more complex algorithms, using deep learning methods. We hope to share more results in coming months! Of course if you have questions, feel free to contact us!

Grégory MAUBON

Grégory MAUBON is Chief Data Officer and digital coordinator at HCS Pharma, a biotech startup focused in high content screening and complex diseases. He manages IT missions and leads digital usages linked to company needs. He is also a Augmented Reality Evangelist (presenter and lecturer) since 2008, where he created and founded in 2010 RA'pro (the augmented reality promotion association). He helped many companies (in several domains) to define precisely their augmented reality needs and supported them in the implementation.


Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.