Not-so-supervised Digital Cytopathology

Background: Digital pathology is booming with numerous deep-learning applications to whole slide image (WSI) analysis. Most studies consider histopathology, pathology of whole tissue sections. In contrast, we focus on cellular-level cytopathology biopsies, containing separate, sparsely located informative groups of cells. These preoperative cytopathology biopsies are signicantly more challenging for diagnosis, often resulting in unnecessary surgeries. We consider automatic prediction of thyroid malignancy from cytopathology WSIs, with the goal of reducing unnecessary surgeries.

Methods: We collected a dataset, > 50 terabytes in size, of 908 (799 training, 109 testing) WSIs; among the rst and largest datasets of cytopathology WSIs. Each WSI includes a binary label (malignant/benign), obtained from post-surgery analysis, and a human diagnostic score per WSI; the dataset also includes an overall 4494 local, patch level, labels annotated by a pathologist. We propose a maximum likelihood estimation framework via a \not-so-supervised" deep-learning algorithm: the WSIs are processed in patches, however, only a tiny number of patches are labeled. Our theoretical analysis establishes a link to multiple- and single-instance learning, leading to an improved training strategy.

Results: The algorithm's performance is compared to 3 cytopathologists using area under the curve (auc) (algorithm: 0:932, experts: 0:931; 0:917; 0:909 ), and average precision (ap) (algorithm: 0:872, experts: 0:848; 0:848; 0:759). By further combining algorithm and human decisions, we reduce indeterminate cases by 55% from 47 to 26.

Conclusion: The proposed algorithm achieves human-level performance and reduces inconclusive diagnoses. Future research includes a multi-center study for evaluating the eect of dierent scanners and staining processes.