Orientation Normalization of Multi-Stain Skin Tissue Cross-Sections

Abstract

Efficient examination of skin tissue specimens is key for pathologists to keep up with an increasing workload. Normalizing the orientation of tissue cross-sections before manual assessment could contribute to a more streamlined digital workflow. In this study, we compare multiple deep learning-based approaches for predicting the rotation angle required to correct the misorientation of skin tissue cross-sections. The models were developed and evaluated using a dataset of 10,649 H&E-stained and 9,731 IHC-stained cross-section images from specimens with melanocytic lesions. Our results show that framing rotation angle prediction as a classification task with the circular target space divided into separate classes performed best, reaching mean absolute errors of 2.77$^\circ$ and 3.56$^\circ$ on the test sets of H&E and IHC-stained cross-sections, respectively, approaching the level of human annotators. Automated orientation normalization, when implemented in whole slide image viewers, could make tissue examination more efficient and convenient for pathologists, while also serving as a valuable preprocessing step for the development of position-aware or multi-stain deep learning models.