Back to all blog posts
Veterinary pathologists play a crucial role in the diagnosis and management of cancer in animals, similar to their human medical counterparts. Canine mast cell tumors (MCTs) are among the most common malignant skin tumors in dogs, and accurate diagnosis and grading are vital for determining prognosis and treatment. In particular, the Ki-67 proliferation marker, a crucial element in cancer grading, can help predict tumor behavior. However, scoring Ki-67 manually is time-consuming and subject to inter-observer variability, leading to inconsistent prognoses and treatment decisions.
This was the key motivation for Dr. Richard Fox, Dr. Melanie Dobromylskyj, and Dr. Annie Ide, veterinary pathologists at UK-based Finn Pathologists, to explore how AI technology could assist in the Ki-67 scoring of canine mast cell tumors. Their focus was on leveraging AI to enhance the accuracy and efficiency of their diagnoses, potentially transforming oncologic decision-making for MCTs. This case study outlines their experience using Aiforia® Create, a leading platform for AI model development in pathology.
Understanding canine mast cell tumors
Canine MCTs range from low-grade tumors that can be managed with surgery alone to high-grade tumors with a poor prognosis, requiring aggressive treatment. Ki-67 scoring is used to measure the rate of cell proliferation within the tumor, an important factor in determining the tumor’s malignancy.
A. Hematoxylin and eosin (H&E) stained section of a high-grade cutaneous mast cell tumor at 30x magnification, demonstrating abundant mitotic figures and moderate cellular atypia.
B. Immunohistochemically stained section of the same mast cell tumor using MIB-1 antibody, highlighting nuclear Ki-67 labeling at 30x magnification.
The pathologists at Finn observed that while Ki-67 was a valuable marker, its manual scoring was labor-intensive and often led to variability among pathologists. This variation could directly affect clinical outcomes, as the decision to pursue chemotherapy or more aggressive treatment depended heavily on this score.
Developing the AI model for Ki-67 scoring
To address these challenges, the team collaborated with Aiforia to develop an AI model to help Ki-67 scoring. Using Aiforia® Create, a versatile tool designed to build and validate deep learning models for image analysis, they embarked on creating an AI model that could not only replicate but improve the speed and consistency compared with manual scoring.
Drs. Fox, Dobromylskyj, and Ide began by defining the specific histopathological features they wanted the AI model to recognize in canine MCTs, with a strong emphasis on tumor cell proliferation as marked by Ki-67. They annotated many digital slides from various cases, ensuring the AI model was trained on a broad spectrum of images and tumor grades.
A typically large training dataset was not needed as the IHC sections were all produced within the lab with two scanners from Finn Pathologists’ archives. The images were scanned at 40X magnification (0.25 μm/pixel), providing high-resolution datasets for the AI model to learn from.
Key features defined for the AI model
The AI model was tasked with learning and analyzing the following key features related to Ki-67 scoring in canine MCTs:
- Segment: Tissue from background
- Segment: Tumor from dermal and subcutaneous tissues
- Segment: Mast cells from non-mast cells (such as lymphocytes, fibroblasts, and eosinophils)
- Detect: Ki-67 positive and Ki-67 negative mast cells
Testing and validating the AI model
Once the initial AI model was trained, it was applied to a test cohort of 30 canine MCT cases. The AI model analyzed the Ki-67 expression and compared its results with those of manual pathologist assessments. The AI model's accuracy was impressive: it closely correlated pathologist scores but overall significantly reduced variability between pathologists.
To further validate the AI model, scoring cut-off, and clinical outcomes, the team applied it to 190 new cases from external veterinary clinics. Again, the AI model demonstrated a high degree of accuracy, and showed a strong significant correlation between tumor recurrence after surgery and overall survival.
Implementing the AI model in clinical practice
The AI model, integrated into Finn Pathologists' diagnostic workflow, has since significantly improved the consistency and speed of Ki-67 scoring by bringing the test in-house. For routine cases, the AI model can now process and provide accurate Ki-67 scores within minutes, allowing veterinary pathologists to focus more on complex diagnostic decisions.
In Phase I, currently in progress, the AI model serves as an assistive tool by highlighting regions with the highest Ki-67 expression using a heatmap. In Phase II, fully automated Ki-67 scoring will be implemented, contingent upon rigorous quality control checks.
More powerful tool for veterinary oncology
The AI-assisted Ki-67 scoring model has not only improved diagnostic accuracy but also enhanced the pathologists’ ability to stratify tumors into low and high-risk groups. This stratification helps guide the treatment plan, ensuring dogs with high-risk MCTs receive more aggressive treatments, while low-risk cases avoid unnecessary interventions.
According to Dr. Fox, "This AI model has transformed how we approach Ki-67 scoring. It’s faster, more reliable, and allows us to provide more consistent and accurate prognoses for our canine patients."
In the future, Dr. Fox suggests that AI-assisted image analysis can be further developed to analyze additional features of MCTs or even other types of veterinary cancers, cytological analysis, and also challenging needle-in-a-haystack tasks like finding rare infectious agents or lesions. "The possibilities are vast," he adds, "and we’re excited to see how AI continues to evolve in veterinary pathology."
Conclusion: AI’s growing role in veterinary pathology
The collaboration between Finn Pathologists and Aiforia highlights the potential of AI to enhance diagnostic precision in veterinary diagnostic pathology. By automating the labor-intensive task of Ki-67 scoring, the AI model will not only help reduce diagnostic time but also improve accuracy, leading to better treatment outcomes for dogs with MCTs.
See the poster about this case: Automated Assessment of Ki-67 Proliferation Index of Canine Mast Tumours Using Computational Deep Learning.