We’ve used computer vision and unsupervised learning algorithm to provide better assistance to doctors and healthcare providers when analyzing Microscopic Images. Technological innovation is making it possible to provide better assistance to doctors and healthcare providers in the diagnosis and treatment of various diseases.
We developed a clustering algorithm to create a consensus map for hand-labeled microscopic images targeting cancer cells, improving diagnosis and providing training data for ML models. Our approach addressed business problems related to modeling complexity, data size, and scalability. The successful POC provided the customer with accurate labeled data for advanced ML supervised models.
Our training base consisted of heterogeneous Microscopic Images, with cancer cells hand labeled by doctors.
Our analytical algorithm for cell labeling consensus used an unsupervised machine learning approach (HDBSCAN) to automatically find optimal clusters. The resulting labels where augmented with data from domain knowledge and hierarchical rules on annotations to create a consensus map.
We’ve tested multiple clustering approaches such as distance-based, hierarchical or density based clustering. These clustering approaches were evaluated based on their ability to accurately identify groups or clusters within the data, as well as their scalability and computational efficiency. We’ve ended up using DBSCAN as it striked the right balance between targeting the correct cells and removing outliers.
Gather Images with Labels
Experiment with Different Clustering Methods
Build Baseline Consensus Map
Add Heuristic Layer