How does AI benefit toxicologic pathology?

Inflammatory bowel disease (IBD) is a complex disease encompassing Crohn’s disease and ulcerative colitis, leading to life-threatening complications and decreased quality of life. A well-known method for rapid candidate compound screening is the dextran sulfate sodium (DSS) colitis model in mice. DSS is administered into the drinking water of mice, resulting in colon ulceration and inflammation, similar to IBD effects. Histopathologic assessment in DSS models often relies mainly on microscopic scoring, a subjective and time-consuming task. 

The researchers’ goal was to examine whether deep learning AI could be used to consistently and quantitatively identify acute inflammation in H&E-stained sections. An AI model was trained using x20 whole slide images of the entire colon to detect key microscopic features in the mouse model of DSS colitis. The detection included the entire colon tissue, the muscle and mucosa layers, and two categories within the mucosa (normal and acute inflammation E1). 

The results obtained using the trained model were consistent with the expected response with the model, suggesting it could correctly segment and identify key microscopic features of the DSS colitis model. With further optimization, the research team sees that this approach could be used as a tool to increase efficiency, provide quantitative data, and decrease variability, subjectivity, and time, as part of screening of candidate compounds for IBD treatment candidates.

Read the full case study →  

Preclinical toxicology and efficacy studies examine all components of possible therapies and their effects on patient bodies. Compounds used for immunotherapy and chemotherapy lie within an extensive list of possible harming factors for hematolymphoid organs such as bone marrow. The importance of bone marrow cells in organ regeneration makes the study of these compounds critical in toxicology safety studies. Evaluating bone marrow cellularity and the ratio of different lineages (e.g., E/M ratio) are always required endpoints. However, routine pathologist evaluation is subjective and does not generate accurate quantitative data.

A study was conducted to evaluate the ability of deep-learning AI models through whole slide image analysis of hematopoietic bone marrow cells from the sternum of cynomolgus macaques. A pathologist developed an easy-to-use AI model alone, training it to enumerate cells in bone marrow. From this, a cell density value was calculated for an objective measure of bone marrow cellularity in tissue sections. 

The AI model’s cell count is comparable to counts by veterinary pathologists, with both groups claiming low error rates. The AI model offers an advantage over the study pathologist by providing objective data, such as the area and cell count of the entire bone marrow section, compared to a subjective severity score provided by a pathologist. 

The AI model is being trained further to recognize different cells within bone marrow for a holistic view of bone marrow cellularity. Additional training will reduce the number of false positives. Similar studies are being conducted in various pathology and toxicology fields to increase efficiency and accuracy throughout research. 

Read the full case study →
See infographic →

Digital tissue image analysis is a powerful computational method for analyzing whole-slide images and extracting large, complex, and quantitative data sets. However, as with any analysis method, the quality of generated results depends on a well-designed quality control system for the entire digital pathology workflow. This requires clear procedural controls, appropriate user training, and the specialists' involvement to oversee key workflow steps. Hence, toxicologic pathologists play a key role in conceiving and implementing a quality control (QC) system. 

Researchers from Charles River Laboratories outline the most common digital tissue image analysis endpoints and potential sources of analysis error. They also recommend approaches for ensuring the quality and correctness of results for both classical and machine-learning-based image analysis solutions. Digital tissue image analysis can increase efficiency, improve consistency, and enable assessments that are not possible or practical via manual evaluation. However, as image analysis technology changes and advances, the need for method qualification and algorithm QC continues regardless of the analysis strategy. 

In selecting an appropriate QC strategy, it is essential to be cognizant of the intended use of the results. The reporting pathologist must understand the methods used for study evaluation, their limitations, the presence of underlying assumptions and resulting biases, and how they can be avoided or minimized. Ultimately, data should only be extracted from samples that completed the QC process and met predetermined performance criteria to ensure high-quality data generation.

Review the publication →

Ovarian follicle differential count is occasionally done in general and reproductive toxicologic studies as long as in other studies, such as animal model characterization for the female genital tract. Counts are time-consuming and difficult. This study aimed to accelerate and standardize image analysis with Aiforia Create. 

The training set included 30 scanned slides uploaded to the cloud-based Aiforia Platform. The AI model was trained in Aiforia Create, and the neural network was compared to annotations done by pathologists. 

The trained convolutional neural network (CNN) delineated with a high level of concordance the three different follicle classes using ground truth established by DR/CS. The AI model identified follicle types without the need for PCNA staining. As a result, AI increased the efficiency of follicle counting by 45x versus manual counting. 

See infographic →