MARTHA - Combining gaze into deep learning for fully quantitative human testicular histology analysis

Despite advances in computational pathology, manual tissue examinations remain the gold standard in diagnostics, resulting in thousands of whole slide image inspections in daily practice. Unfortunately, examination strategies and identified regions are often lost after inspection, preventing the use of manual efforts to train deep learning algorithms. To address this, we introduce MARTHA, a tool that combines passive eye tracking - which captures pathologists' gaze during tissue examinations - with deep learning-based image analysis. This approach simplifies and accelerates data interaction while improving the accuracy and efficiency of computer-assisted diagnostics by incorporating expert attention into neural networks. An intuitive graphical interface integrates traditional interaction methods, allowing seamless use across clinical routines and biomedical image domains. We evaluated MARTHA on human testicular tissues, generating the largest annotated dataset for the human testis, with over 83,000 cell nuclei from approximately 8000 tubules. Our results demonstrate strong performance in data interaction efficiency and semantic segmentation, offering valuable insights into testicular phenotypes and supporting pathologists in enhancing their analyses, which is a crucial step towards improving the diagnosis and treatment of infertile men with unfulfilled child-wish.