Deep Learning Discovers New Morphological Features while Predicting Genetic Alterations from Histopathology of Papillary Thyroid Carcinoma

<b><i>Background:</i></b> Papillary thyroid carcinoma (PTC) is the most common malignant tumor of the endocrine system. <i>BRAF</i> mutations occur in 40-60%, <i>panRAS</i> mutations in 10-15%, and different gene fusion events such as <i>RET</i> fusions in 7-35% of these neoplasms. Artificial intelligence (AI) methods could be used to predict genetic changes from conventional histopathological slides. <b><i>Methods:</i></b> In this retrospective study, we used two independent cohorts of patients with PTC, totaling 662 cases for the establishment of our AI pipeline. The Cancer Genome Atlas cohort (496 cases) served as the developmental cohort, while the Mainz cohort (166 cases) served as an independent external test cohort. <i>BRAF</i>, <i>panRAS</i>, and fusion status was determined for all of these patients as target variables. Vision Transformer was trained on digitized annotated hematoxylin and eosin-stained slides for the presence of these alterations. Highest probability image tiles were used to identify new morphological criteria associated with the genetic changes. <b><i>Results:</i></b> The trained model resulted in an area under the receiver operating characteristic curve of 0.882 (confidence interval 0.829-0.931) for <i>BRAF</i>, 0.876 (0.822-0.927) for <i>panRAS</i>, and 0.858 (0.801-0.912) for gene fusions. Accuracy was 79.3% (72.7-85.8%) for <i>BRAF</i>, 89.3% (84.2-94.0%) for <i>panRAS</i>, and 84.7% (78.8-90.2%) for gene fusions. The performance on the validation set was almost identical to that on the test set. Analyzing the highest predictive tiles, novel morphological criteria for fusion-associated PTC could be discovered. <b><i>Conclusions:</i></b> Our study demonstrates that predicting genetic alterations in digitized histopathological slides using AI is feasible in patients with PTC. Our model showed high accuracy in predicting these changes, making it potentially suitable for pre-screening. Explainability approaches uncovered previously undescribed morphological patterns associated with certain genotypes. Providing pathologists with these AI-based features could improve their accuracy. Assuming further positive prospective validation, this discovery could contribute to a deeper understanding of PTC.