Pathology specimen analysis ～Abnormality screening and virtual staining～

Challenge

Pathological examinations in drug safety assessments are traditionally performed by direct microscopic observation. However, this process is time and labour intensive because toxicologists must thoroughly examine the entire surface of every slice generated.

In addition, as the diversification of drug discovery modalities increases, so does the likelihood of encountering new types of lesions that have not been previously observed. This raises concerns that AI models trained solely on the appearance of known lesions will be limited in practical implementation.

Furthermore, the effects of artifacts generated during the specimen preparation process and inter-facility variations must also be taken into consideration.

LPIXEL’s Solution – AI Technology and Impact

AI-based pathological specimen screening

Two specialized AI models are trained on a wide range of normal kidney tissue images, enabling them to detect deviations from normal (abnormalities) in new specimens:

• Detection of tubular abnormalities visible at lower magnification (capturing disruption of the overall tissue structure)
• Detection of glomerular abnormalities (Identifying minute changes)

By combining the outputs of these two dedicated models, the AI system determines whether a specimen is overall normal or abnormal.

As this approach evaluates abnormality based on the degree of deviation from normal tissue, rather than relying solely on recognizing specific, predefined lesions, this method is expected to be effective even for previously unseen, novel toxicities.

We have successfully achieved detection of both tubular and glomerular lesions, along with visualizations that illustrate the extent of deviation from normal tissue structure.

Virtual Staining

By generating virtually stained images from unstained samples using AI, inter-facility variability in staining can be eliminated to achieve consistent, uniform image analysis. This approach also allows (virtual) staining to be applied to tissue specimens reserved for genetic analysis without degrading the original material.

As another example, by generating immunostained images from existing HE slides, researchers can investigate specific biological mechanisms by leveraging the large archive of HE images already available. This also allows researchers to pinpoint regions of interest before performing the actual immunostaining.