AI Transforms Pathology
Source: nature.com
AI in Pathology
Pathologists are facing increasing demands, with shortages occurring in many countries. Their work includes sectioning and staining tissues, viewing them under a microscope, and conducting tests for genes and molecular markers.
Artificial intelligence (AI) may offer a solution, with tools that can highlight suspicious regions, standardize diagnoses, and find patterns not visible to humans. According to Bo Wang, AI has the potential to improve diagnostic accuracy, reproducibility, and efficiency, and to enable new research by mining pathological and molecular data.
Developments in AI Models
Slides have been digitized, enabling pathologists to study samples on-screen, and these images have been used to develop AI assistants. The success of AI chatbots has led to applying similar techniques to pathology. AI models are being designed to classify illnesses, predict treatment outcomes, and identify biological markers. Some models can mimic the entire pathology process, from slide analysis to report writing, according to Faisal Mahmood.
Some researchers are cautious, saying that AI models need further validation and that the nature of some models poses deployment challenges. Hamid Tizhoosh notes the need for reliable, accurate, and robust results when these tools are used in hospitals.
Foundation Models
Early AI tools for pathology were designed for specific tasks, such as detecting cancer in breast-tissue samples. Foundation models, which can adapt to a broad range of applications, offer an alternative. Large language models that drive generative-AI tools like ChatGPT are among the best-known foundation models; however, pathologists lack a resource as vast as the Internet to train their software.
In 2023, Meta released DINOv2, a foundation model for visual tasks. Mahmood and his team launched UNI2 in March 2024, a general-purpose model for pathology, using a data set of over 100 million images from 100,000 slides. The model outperformed existing computational-pathology models on classification tasks. UNI 2 has an expanded training data set.
Another foundation model, CONCH, used diverse data sets and images from pathology slides and text from medical databases. CONCH could perform classification tasks better than other models and could classify and caption images. UNI and CONCH are available on Hugging Face.
Researchers have used them for applications such as grading and subtyping neural tumours, predicting treatment outcomes, and identifying gene-expression biomarkers. Mahmood stated surprise at the level of interest in computational pathology, with over 1.5 million downloads and hundreds of citations.
Other groups have developed their own foundation models. Microsoft’s GigaPath is trained on over 170,000 slides to do tasks such as cancer subtyping. mSTAR, designed by Hao Chen and his team, combines gene-expression profiles, images, and text. It was designed to detect metastases and subtype cancers and is also available on Hugging Face.
AI Copilots and Concerns
Mahmood and Chen’s teams have built ‘copilots’ based on their models. Mahmood’s team released PathChat in June 2024, a generalist AI assistant, fine-tuned with information from PubMed, case reports, and other sources. Pathologists can use it to have conversations about images and generate reports. Chen’s team developed SmartPath, a chatbot being tested in hospitals in China for assessments of breast, lung, and colon cancers.
PathChat and SmartPath can plan, make decisions, and act autonomously. PathChat can streamline workflows by highlighting likely positive cases, ordering tests, and writing reports. Jakob Kather believes foundation models are transformative but await regulatory approval, estimating two to three years until widespread availability. Anant Madabhushi notes accuracy as a key concern, stating that cross-validation can produce optimistic results and that external validation is best.
Tizhoosh and colleagues tested pathology foundation models, including UNI and GigaPath, using a zero-shot approach with data from the Cancer Genome Atlas. The assessed models were, on average, less accurate at identifying cancers than a coin toss would be, though some performed better for specific organs.