Tumor Segmentation
Project Details
Problem: Analyzing gigapixel whole slide images is time-consuming for pathologists, though absolutely critical to the diagnosis of cancer. If cancerous regions can be identified faster for review by a pathologist, the overall analysis can be completed more efficiently. Locating cancerous regions is complicated by the varying appearance of different grades.
Solution: This project involved guiding a junior ML team on best practices for segmenting tissue types. We initially found a number of inconsistencies in the pixel-level annotations created by pathologists, so we started by refining and standardizing the annotation process. As the labeled dataset became cleaner, we were able to iterate more effectively on the segmentation model.