The prognostic significance of tall cells and lymphocytes in papillary thyroid carcinoma : the use of deep learning algorithms

Papillary thyroid carcinoma (PTC) is the most common thyroid cancer variant with an overall excellent prognosis. However, some histological subtypes demonstrate a more aggressive disease progression and thus require more attention from clinicians. The tall cell subtype of papillary thyroid carcinoma (TC-PTC) is one such subtype; it is characterized by the presence of tall epithelial cells, comprising approximately 30% of the tumor volume. These cells are three times taller than they are wide, and display nuclear features consistent with classical PTC. This definition, however, is challenging to adhere to by traditional microscopy, which results in large inter-observer variability between pathologists when diagnosing TC-PTC. The aim of this thesis was to study the tall cell (TC) threshold needed for an adverse outcome. We trained a deep learning (DL)-based algorithm for tall cell detection and scoring. The TC-algorithm detected TC areas with a sensitivity and specificity of 93.7% and 94.5%, respectively, and non-TC areas with a sensitivity and specificity of 90.3% and 94.1%, respectively, in the test set. The performance of the TC-algorithm was compared to visual TC scoring on an internal validation dataset. A higher TC score assessed by the TC-algorithm correlated with a reduced relapse-free survival (RFS) for 10%, 20%, and 30% TC thresholds. The visually assessed TC scores did not predict survival at any of the analyzed TC thresholds. The trained TC-algorithm was further externally validated using held-out multicenter PTC datasets, one originating from Auria Biobank, Turku, Finland, and the other from the University of Bern, Switzerland. In the external validation, the DL-based algorithm detected TC areas with a sensitivity and specificity of 90.6% and 88.5%, respectively, while the non-TC areas were detected with a sensitivity and specificity of 81.6% and 92.9%, respectively. In the external validation datasets, a higher TC score correlated with a reduction in relapse-free survival using a 20% and 30% TC threshold. Immune cells of the tumor microenvironment play an important role in the development and progression of cancers and may have either tumor promoting or suppressing effects. High numbers of tumor-infiltrating lymphocytes (TILs) have been associated with a favorable outcome in certain cancers such as breast and colon cancer, and have also been shown to correlate with a favorable prognosis in PTC. Quantifying TILs is routinely performed by visual evaluation and estimation using a traditional microscope, which is time-consuming and subject to inter- and intra-observer variability. In this thesis, we also trained a DL-based algorithm for segmenting TIL areas in PTC. We trained this model using a novel antibody-supervised learning approach with a pan-leukocyte CD45 antibody staining as ground truth, and applied the model to hematoxylin and eosin (HE)-stained tissue slides. Twelve PTC whole slide images (WSIs) were analyzed by the trained algorithm, which had an intersection over union of 0.82 for detecting TIL areas in HE-stained tissue slides when comparing the algorithm predictions to the ground truth anti-CD45 mask. Conclusively, the findings suggest that a DL-based algorithm approach can register and find TCs with high sensitivity and specificity, even in externally collected, independent datasets without any supportive training. An algorithm TC threshold of 30% correlated with a reduction in relapse-free survival, and is suggested to be used when diagnosing TC-PTC. All cases with a TC score above 10%, i.e., PTC with tall cell features, should be reported by the pathologist. We also conclude that the proposed method of generating antibody-supervised annotations using destain-restain immunohistochemistry-guided annotations resulted in highly accurate segmentation of TIL-rich areas in HE-stained tissue images.