1623-P: Artificial Intelligence with QuPATH in Pancreatic Cell Classification

• Published in: Diabetes (New York, N.Y.) Vol. 73; p. 1
• Main Authors: Guyot, Michael, Williams, MacKenzie D, Bumgarner, Benjamin M, McGrail, Kieran M, Campbell-Thompson, Martha, Haller, Michael J, Bruggeman, Brittany S
• Format: Journal Article
• Language: English
• Published: New York American Diabetes Association 01-06-2024
• Subjects: Accuracy; Algorithms; Artificial intelligence; Cell size; Datasets; Diabetes; Diabetes mellitus (insulin dependent); Immunofluorescence; Information processing; Neural networks; Pancreas
• ISSN: 0012-1797; 1939-327X
• DOI: 10.2337/db24-1623-P

We developed Artificial Intelligence classification algorithms in QuPATH containing both supervised and unsupervised components to quantify and classify the pancreas cell types into their respective subtypes based on immunofluorescence (IF) data. The staining panel was performed in formalin-fixed paraffin embedded (FFPE) pancreatic tissue sections from a pancreas organ donor with no diabetes provided by nPOD. From the IF images, a training region was selected. This region was then segmented using a watershed algorithm and then an expert classified each cell in the region. In total, there were 6,899 cells, with an emphasis on the acinar and ductal cell types. This training data was used to train a Random Trees (RT), K Nearest Neighbor (KNN) and Artificial Neural Network (ANN) algorithm. Another region was selected for a validation data set, and once again, each cell was manually classified by an expert in the field. The accuracy of each algorithm was then assessed on the validation set. The accuracy of each algorithm is as follows: RT (89.53%), KNN (85.55%) and ANN (82.97%). In the future, we will include multiple no diabetes pancreases as well as pancreases from the different stages of T1D progression to ensure diversity of the training and validation data. With a more diverse training data set, we predict that ANN's will become more accurate, and could potentially prove to be more useful than the RT algorithm. With a more diverse validation data set, we can look at the possibility of overfitting artificially inflating the accuracy of our algorithms.From this, we plan to extract relevant feature information such as cell size, number and density among others. After we have the architecture solidified, we hope to use it to analyze differences in control and diseased pancreases, specifically with regards to Type 1 Diabetes (T1D). We hope this will uncover previously overlooked information that may prove critical to the understanding of the mechanism of disease progression of T1D.