Embedded Files
This website is no longer supported. Please visit my new site at vdanilov.com for the latest updates and projects
Deep learning to detect and count cells in microscopy images
Deep learning to detect and count cells in microscopy images
Customer: Volastra Therapeutics (New York, United States)
Summary: The aim of this study is to prevent the spread of cancer by stopping it at its onset. To achieve this, a pipeline has been developed to locate and count different types of objects, such as primary nuclei, micronuclei, mitosis, and apoptosis. The ratio between the number of these objects can indicate how well the cancer treatment is working. The proposed solution for counting and locating both primary nuclei and micronuclei is based on an object detection approach. The EfficientDet network, which uses a pre-trained backbone on the ImageNet dataset and a BiFPN feature network, was used as the main solution. The dataset used for this study was gathered from internal Volastra sources, the Image Data Resource, and the HMS LINCS database, and included 36 cell lines with 1664 images (1004 for training, 460 for validation, and 200 for testing). The best network achieved a mean average precision of 66% for micronuclei and 85% for primary nuclei.
Collaborators: Oleg Talalov (Amazon, Vancouver, Canada), Christina Eng (Volastra Therapeutics, New York, United States), Akanksha Verma (Volastra Therapeutics, New York, United States)
Project type: Commercial
Figure 1. EfficientDet network architecture used in this study
Tiling with regular window shift
Tiling with padding
Tiling with unfixed-size window
Figure 2. Representation of different overlap-tile strategies used for the model development
Figure 3. Testing model on the MB-231 cell line (Magnification: 63X)
Figure 4. Testing model on the HeLa cell line (Magnification: 40X)
Figure 5. Testing model on the CAL51 cell line (Magnification: 20X)
Page updated
Google Sites
Report abuse