Automated Tissue Analysis

Automated Tissue Analysis:

Pathologists spend much of their day viewing tissue slides and identifying features of the tissue that imply the presence of disease. There is significant interest in applying ML to digital pathology because ML may automate parts of the diagnostic process, lowering cost and increasing availability of diagnosis. This project attempts to automate one task: identifying and counting eosinophil cells in digitized nasal tissue slides. The eosinophil cell density maps to one of three diagnoses: mild, moderate, or severe Chronic Sinusitis, and the diagnosis guides treatment. Currently, a pathologist diagnoses this condition manually.

Working with the Cooperative Human Tissue Network and the Pathology Department of the University of Pennsylvania, under the leadership of Dr. Virginia Levolsi, a pathologist specializing in nasal tissue, an initial system for tissue analysis has been implemented and tested. This work shows the approach to be promising; however, the computation time is excessive and the reported accuracy needs improvement. The system breaks down slides into individual cells which are then classified using a convolutional neural network that was trained on data provided by Dr. Levolsi which contained 6000 cells. The system is 55% accurate on 48 slides and requires about 2-3 hours to process a single slide. The proposed project would explore techniques to improve accuracy and reduce processing time.