Conventional stain and AI-derived pathology tissue
Safety Assessment
Regina Kelder

Deep Learning AI Models: A New Tool for Pathologists (Art of the Science)

How AI models are helping pathologists spot abnormalities efficiently and more consistently

What are we looking at in these images above?

The image on the left represents a traditional stained slide for a pathologist and on the right a Deep Learning AI analyzed section with colored masks that highlight different compartments of the tissue and whether abnormalities may be present. Charles River Laboratories is developing these AI models to provide decision support for pathologists in their research supporting the development of medications for important veterinary and human diseases.

How did you generate the AI-image?

The skin sections were imaged on a digital scanner at 40x magnification and uploaded to a partner’s (Deciphex) cloud for training. Images were viewed on a digital study browser and annotations made by the investigators to teach the computer the normal and abnormal areas on the scanned slides. The annotations were used to develop an AI-based convolutional neural network model that could reliably predict normal and abnormal skin regions. The model paints colored masks onto the digital image for the pathologists that define the different normal and abnormal regions and can be rapidly viewed alongside the original digitized slides.

How long have you been using the AI models and why are they better than more conventional methods?

Traditional machine learning AI has been used at CRL for almost two decades. The limitations for traditional approaches is that the human has to find a way to show the computer each feature that could be used by the computer for region/cell identification, which is very difficult and time consuming (and sometimes impossible). However, deep learning based AI is a relatively new and more powerful tool that allows the investigator just to show the computer general areas that are important to differentiate and lets the computer decide on the features to use to make the decision between those areas. We have been working with deep learning AI for about three years and have really accelerated its use in the last 18 months with our partners from Deciphex, Ltd. We have observed marked increases in efficiency and improved consistency using Deep Learning AI vs. traditional microscopy techniques.

Tell us something cool about the Digital Pathology team?

The digital pathology team is a global, fully remote team of pathologists and scientists from across the world (Australia to Oregon). We have 4 pathologists and 2 co-ops on the team and partner with over 20 scientists globally at CRL, professors and interns/externs at veterinary school pathology and biomedical engineering departments, and partners at Deciphex, Ltd as well as vendors like Aiforia and Visopharm. We are quite the nerdy group and really want digital pathology to advance our science.