Development of a Deep Learning Tool to Support the Assessment of Thyroid Follicular Cell Hypertrophy in the Rat.

Thyroid tissue is sensitive to the effects of endocrine disrupting substances, and this represents a significant health concern. Histopathological analysis of tissue sections of the rat thyroid gland remains the gold standard for the evaluation for agrochemical effects on the thyroid. However, there is a high degree of variability in the appearance of the rat thyroid gland, and toxicologic pathologists often struggle to decide on and consistently apply a threshold for recording low-grade thyroid follicular hypertrophy. This research project developed a deep learning image analysis solution that provides a quantitative score based on the morphological measurements of individual follicles that can be integrated into the standard pathology workflow. To achieve this, a U-Net convolutional deep learning neural network was used that not just identifies the various tissue components but also delineates individual follicles. Further steps to process the raw individual follicle data were developed using empirical models optimized to produce thyroid activity scores that were shown to be superior to the mean epithelial area approach when compared with pathologists' scores. These scores can be used for pathologist decision support using appropriate statistical methods to assess the presence or absence of low-grade thyroid hypertrophy at the group level.