From Microscopes to Machine Learning:
Identifying and Counting Blood Cells with Deep-Learning Models
Download Full White Paper
Our client is a medical diagnostics company developing a point-of-care (POC) blood cell analyzer, which utilizes deep-UV microscopy to create biomolecular maps for a complete blood count (CBC) with a 5-part white blood cell (WBC) differential. The device is designed to serve oncology patients undergoing chemotherapy at risk of neutropenia, thrombocytopenia, and anemia.
To ensure timely treatment, these patients require frequent CBC testing to determine whether they are eligible for chemotherapy or need a blood transfusion and, if they are febrile, whether they must be admitted to the hospital.
Preliminary work has demonstrated the ability to use machine learning algorithms, specifically deep neural networks (DNNs), to perform a 5-part WBC differential. We have expanded that automation capability to include red blood cells (RBCs) and WBCs with 95% or greater accuracy.
Challenges
The preliminary work to create the 5-part WBC differential showed an accuracy of 94% across all classes. To achieve this accuracy, two segmentation DNNs, two classification DNNs, and two traditional computer vision (CV) techniques were applied to the data. Furthermore, the data were entirely hand-labeled for each of the DNNs, which was time-consuming.
The primary challenges were twofold. Firstly, we had to curate and modify existing datasets to meet our needs. Secondly, we had to train and validate new algorithms to ensure they met the additional requirements of covering the expanded feature space.
Expertise Employed
- PyTorch
- TensorFlow
- OpenCV
- YOLO
- Segmentation transformer (SegFormer)
- RESNET
- Deep learning
- Artificial Intelligence (AI)
- Deep neural network
- Data bootstrapping a large, labeled dataset from an initial set with limited labeling
- Transfer learning
- Hematology
- Medical image processing
