Automatic assessment of AgNOR-scores in canine cutaneous mast cell tumors using a Deep Learning-based algorithm

Introduction Argyrophilic nucleolar organizing regions (AgNORs) are nucleolar substructures involved in ribosomal RNA transcription. It has been reported, that the number of AgNORs per nucleus correlates with the cell proliferation rate and that the average number of AgNORs per cell (AgNOR-score) is of prognostic value for survival prediction. Enumeration of AgNORs is a tedious task, which can be accelerated by algorithmic methods.

Material and Methods We created a ground truth dataset consisting of 29 images (1,569 x 1177 Pixel each) with 23,036 cell annotations. For ten images, AgNOR-scores were reevaluated by six pathology experts. An ensemble of five detection models was trained on five random subsets of training data. The mean AgNOR-score per image of the ensemble was measured and compared to the mean AgNOR-scores of the experts.

Results We found a mean-squared-error of 0.054 and a mean-absolute-error of 0.201 between the AgNOR-scores of the model ensemble and the mean AgNOR-scores of the experts.

Conclusion The small error between the values of the ensemble and those of the experts showed that deep learning is capable of automatic evaluation of AgNOR. These findings enable large-scale, computer-aided studies investigating the prognostic impact of AgNOR on biological tumor behavior.

Publikationsverlauf

Artikel online veröffentlicht:
11. August 2023