Abstract
In recent years, there has been growing expectation for the practical implementation of AI in the field of toxicologic pathology to alleviate the burden on pathologists reading a lot of tissue slides.
Fujifilm is conducting research on AI for screening of a lot of tissue slides. We have confirmed that a model created without training on real lesion images can detect 12 types of rat-liver lesions based on deviations from normal morphological characteristics.
Building a dataset with a sufficient scale and quality is crucial for the successful implementation of AI. Particularly, in toxicologic pathology, there exist challenges in both (1) collecting tissue images and (2) annotating them. (1) It is difficult to obtain lesion images, especially for rare lesions. Additionally, there are morphological variations in normal tissues due to factors such as sex and aging. Furthermore, technical variations in slide preparation are reflected in images, so there are various attributes to consider in collecting tissue images. (2) Lesions that are spatially and morphologically ambiguous introduce subjectivity in the annotation process. Ideally, the involvement of multiple pathologists would enhance the reliability of annotations, but this is often impractical due to their demanding schedules. Consequently, the current implementation is functionally restricted to in-house slides and specific lesions.
To implement a versatile AI, it is important to ensure diversity in images and reduce the burden of annotation for each institution. Cross-institution programs will accelerate these efforts.
