Artificial intelligence (AI) has made significant strides in visual analysis, particularly in recognizing patterns, which is crucial for detecting cancer. Radiologists use X-rays and MRIs to identify tumors, while pathologists examine tissue samples under a microscope to find specific patterns. These patterns can indicate the severity of cancer, the potential effectiveness of treatments, and the likelihood of cancer spreading.
AI could theoretically be a great help in this process. According to Andrew Norgan, a pathologist and medical director at the Mayo Clinic, “Our task is pattern recognition. We examine the slides and gather information that has proven important.” The goal is to develop AI models that can enhance the accuracy and speed of cancer diagnosis.
There have been several attempts to create effective AI models for cancer detection. One such model, called Atlas, was developed by Aignostics and the Mayo Clinic. Atlas was trained on 1.2 million tissue samples from 490,000 cancer cases. It was tested against six other leading AI pathology models and outperformed them in six out of nine tests. For instance, Atlas matched human pathologists’ conclusions in 97.1% of colorectal cancer cases. However, its performance was less impressive in classifying prostate cancer biopsies, with only a 70.5% accuracy rate.
The best way to understand cancer cells in tissue is still through examination by human pathologists. Current AI models are measured against human performance. While some models perform well in specific tasks, they often fall short in others. For AI models to be clinically useful, they need to be exceptionally accurate.
One major challenge is the lack of digital data. In the U.S., less than ten percent of pathology practices are digitalized, meaning many tissue samples are not available in a digital format for AI training. Although European practices are more digitalized, there is still a scarcity of comprehensive datasets for AI training. Without diverse datasets, AI struggles to recognize a wide range of anomalies, including rare diseases.
To address this, the Mayo Clinic began digitalizing its pathology practices and archived 12 million slides. They used a robot to take high-resolution images of these samples, creating a dataset of 1.2 million samples for AI training.
Another issue is the size of digitalized tissue samples. Biopsy samples, though small, are magnified to create digital images containing billions of pixels. This requires significant storage and careful selection of image parts for AI training. The Mayo Clinic used a “tiling” method to create multiple snapshots of the same sample for AI input.
Determining the right benchmarks for AI models is also crucial. Atlas was tested in molecular benchmarks, which involve finding clues in tissue images to predict molecular-level events. For example, certain genes are crucial in cancer because they repair DNA replication errors. AI could potentially predict molecular changes that are not visible to humans.
Currently, AI models, including Atlas, have limited success in these molecular tests. Atlas achieved an average score of 44.9% in molecular tests, indicating room for improvement. According to Carlo Bifulco, Chief Medical Officer at Providence Genomics, while Atlas shows progress, significant advancements require new models and larger datasets.