AI-Based Histopathological Lesion Assessment for Safety Evaluation in Rat Kidney and Testis

<p>In non-clinical safety studies at the screening stage, microscopic examination of numerous histopathological specimens is conducted. Pathological examination assisted by artificial intelligence (AI) is promising; however, challenges remain in detecting unknown lesions and addressing inter-facility variability in specimen preparation. This study focuses on AI-based lesion visualization and quantitative assessment in renal and testicular pathology.For the kidney, we developed an anomaly detection model for tubular and glomerular pathology using only normal images for training. For the testis, we built deep learning models for stage classification of seminiferous tubules and cell-type classification of spermatogenic cells. While pathologists typically perform qualitative assessments, AI enables quantitative evaluation, enhancing detection of subtle lesions.The renal tubular model achieved 100% sensitivity and specificity in specimen-level classification. The glomerular model reached 88% sensitivity and 53% specificity. In testis analysis, the AI models achieved 96% accuracy for tubule staging and 98% for cell-type classification, confirming practical applicability. The model was evaluated using multi-source specimens, confirming its generalizability.Future work includes integrating AI into safety studies and expanding its application to other organs.</p>