New artificial intelligence tools for cancer treatment

The new AI system announced on September 4th: Nature, The researchers said their technique is a step ahead of many current AI approaches in cancer diagnosis. (DOI 10.1038/s41586-024-07894-z)

Current AI systems are typically trained to perform specific tasks, such as detecting the presence of cancer or predicting a tumor's genetic profile, and tend to work only on a small number of cancer types. In contrast, the new models can perform a wide range of tasks, have been tested on 19 cancer types, and have the flexibility of large-scale language models such as ChatGPT.

While other foundational AI models for medical diagnosis based on pathology images have emerged recently, this is believed to be the first to predict patient outcomes and validate it across multiple international patient groups.

“Our goal was to create an agile, versatile ChatGPT-like AI platform capable of performing a wide range of cancer assessment tasks,” said Cunxin Yu, assistant professor of biomedical informatics at the Blavatnik Institute of Harvard Medical School and senior author of the study. “Our model proved highly useful in multiple tasks related to cancer detection, prognosis, and treatment response across a range of cancers.”

The AI ​​model reads digital slides of tumor tissue, detects cancer cells based on cellular features seen in the images, and predicts the tumor's molecular profile, with accuracy superior to most current AI systems. The model can predict patient survival across multiple cancer types and accurately identify features of the tissue surrounding the tumor (also known as the tumor microenvironment) that are associated with a patient's response to standard treatments such as surgery, chemotherapy, radiation therapy and immunotherapy. Finally, the tool appears capable of identifying specific tumor characteristics not previously known to be associated with patient survival, generating new insights, the team said.

The researchers say the findings add to evidence that AI-powered approaches can improve clinicians' ability to assess cancer efficiently and accurately, including by identifying patients who are likely to respond poorly to standard cancer treatments.

“If further validated and widely deployed, our approach, and approaches like it, could potentially identify early cancer patients who may benefit from therapies targeted at specific molecular mutations – a capability that is not uniformly available around the world,” Yu said.

Training and Performance

The team's latest research builds on Yu's previous work on AI systems for the assessment of colon cancer and brain tumors, which demonstrated the feasibility of the approach within specific cancer types and specific tasks.

The new model, called CHIEF (Clinical Histopathology Imaging Evaluation Foundation), was trained on 15 million unlabeled images segmented into regions of interest. It was then further trained on 60,000 whole-slide images of tissues including lung, breast, prostate, colon, stomach, esophagus, kidney, brain, liver, thyroid, pancreas, cervix, uterus, ovaries, testes, skin, soft tissue, adrenal glands, and bladder. By training the model to look at both specific parts of an image and the whole picture, it is now able to relate specific changes in one area to the overall situation. According to the researchers, this approach allows CHIEF to interpret the image more holistically, taking into account the broader context rather than just focusing on a specific area.

After training, the team tested CHIEF's performance on more than 19,400 whole-slide images from 32 independent datasets collected from 24 hospitals and patient cohorts around the world.

Overall, CHIEF outperformed other state-of-the-art AI methods by up to 36 percent on tasks such as detecting cancer cells, identifying tumor origins, predicting patient outcomes, and identifying the presence of genes and DNA patterns associated with treatment response. Because of its generic training, CHIEF performed equally well regardless of how the tumor cells were collected – via biopsy or surgical resection – and was equally accurate regardless of the technique used to digitize the cancer cell samples. This adaptability, according to the researchers, allows CHIEF to be used in a variety of clinical settings, marking an important step beyond current models, which tend to only perform well when reading tissue collected with specific techniques.

Cancer Detection

Across 15 datasets covering 11 cancer types, CHIEF achieved nearly 94 percent accuracy in cancer detection, significantly outperforming current AI approaches. In five biopsy datasets collected from independent cohorts, CHIEF achieved 96 percent accuracy across multiple cancer types, including esophageal, gastric, colon, and prostate. When researchers tested CHIEF on never-before-seen slides of surgically removed tumors of the colon, lung, breast, endometrium, and cervix, the model achieved more than 90 percent accuracy.

Predicting the molecular profile of tumors

A tumor's genetic makeup provides important clues for determining the tumor's future behavior and optimal treatment. To get this information, oncologists order DNA sequencing of tumor samples, but deep genomic profiling of cancer tissue is not routinely or uniformly performed around the world due to the cost and time involved in sending samples to specialized DNA sequencing laboratories. Even in well-resourced areas, the process can take weeks. This is a gap that AI can fill, Yu says.

Rapidly identifying cellular patterns on images that suggest specific genomic abnormalities could be a rapid, cost-effective alternative to genome sequencing, the researchers said.

CHIEF outperformed current AI methods that look at microscope slides and predict genomic mutations in tumors. The new AI method successfully identified features associated with several key genes associated with cancer growth and suppression, and predicted key genetic mutations associated with tumor response to various standard of care treatments. CHIEF also detected specific DNA patterns associated with the degree to which colon tumors respond to an immunotherapy called immune checkpoint inhibition. Looking at whole tissue images, CHIEF identified mutations in 54 commonly mutated cancer genes with an overall accuracy of over 70%, outperforming current state-of-the-art AI methods for genomic cancer prediction. Its accuracy was even higher for specific genes in specific cancer types.

The team also tested CHIEF's ability to predict mutations associated with response to FDA-approved targeted therapies across 18 genes across 15 anatomical sites. CHIEF achieved high accuracy across multiple cancer types, including 96 percent accuracy in detecting mutations in a gene called EZH2, which is common in a type of blood cancer called diffuse large B-cell lymphoma; 89 percent accuracy in detecting BRAF mutations in thyroid cancer; and 91 percent accuracy in detecting NTRK1 mutations in head and neck cancer.

Predicting patient survival

CHIEF accurately predicted patient survival based on tumor histopathology images obtained at the time of initial diagnosis. Across all cancer types and patient groups studied, CHIEF distinguished between long-term and short-term survivors. CHIEF outperformed other models by 8 percent. And, in patients with more advanced cancers, CHIEF outperformed other AI models by 10 percent. Overall, CHIEF's ability to predict high or low risk of death was tested and confirmed in patient samples from 17 different institutions.

Extracting new knowledge about tumor behavior

The model identified distinctive patterns on the images that were associated with tumor aggressiveness and patient survival. To visualize these regions of interest, CHIEF generated heat maps on the images. When human pathologists analyzed these AI-derived hotspots, they found intriguing signals that reflect interactions between cancer cells and surrounding tissues. One such feature was the presence of more immune cells in the tumor areas of long-term survivors compared to short-term survivors. This finding makes sense, Yu noted, because the presence of more immune cells could indicate that the immune system has been activated to attack the tumor.

When looking at tumors from patients with short survival, CHIEF identified regions of interest characterized by abnormal size ratios between different cellular components, more atypical features in cell nuclei, weaker connections between cells, and less presence of connective tissue around the tumor. These tumors also had more dying cells around them. For example, in breast tumors, CHIEF identified the presence of necrosis, or cell death, in the tissue as a region of interest. Conversely, breast cancers with high survival rates were more likely to have preserved cellular structures similar to healthy tissue. The visual features and regions of interest associated with survival differed depending on the type of cancer, the team noted.

Next steps

The researchers said they plan to improve CHIEF's performance and enhance its capabilities in the following ways:

• Providing additional training in tissue imaging for rare and non-cancerous diseases
• Contains samples of precancerous tissue before cells become fully cancerous
• By exposing the model to more molecular data, we can improve its ability to distinguish between cancers with different levels of aggressiveness.
• Train a model to predict the benefits and side effects of novel cancer treatments in addition to standard treatments