An Evaluation of Cell-Type Specific Artificial Intelligence Algorithms in Adenocarcinoma in Situ in Papanicolaou Tests

Summary: Introduction: Diagnosis of adenocarcinoma in situ (AIS) on Papanicolaou (Pap) tests is challenging for pathologists. Artificial intelligence (AI)-based Pap test screening has been extensively studied, however, less work has been done on glandular lesions. Materials and Methods: Seventy Pap test case images from the Hologic digital pathology education website were selected including AIS cases, benign normal (NILM) cases, atypical glandular cell (AGC) cases and adenocarcinoma (AdenoCA) cases. Different cell types were manually annotated and then analyzed using an open software platform (QuPath) to establish an algorithm to detect the different cell types. Thirty randomized cases from the Hologic website were then used to define the percentage of AIS, AGC, AdenoCA under the annotated areas on different cell types. Results: The QuPath algorithm defined an average percentage of AIS under the annotated areas as 55.64% in the AIS case group, 9.85% in the NILM group, 29.01% in AGC group, and 14.00% in adenocarcinoma group. The area under the ROC curve of this algorithm is 0.847 which indicates that the algorithm classifies two randomly chosen test cell clusters correctly as AIS or non-AIS with a probability 84.7%. Conclusions: Using images from the Hologic digital pathology website that reflect the panel images presented to cytologists and pathologists, we were able to independently create an algorithm that could classify the AIS cases with good accuracy.