Models similar to ChatGPT can diagnose cancer
Guide treatment choices and predict survival across multiple cancer types
Scientists at Harvard Medical School have designed a multifunctional AI model similar to ChatGPT that can perform a range of diagnostic tasks for a variety of cancers.
Researchers say the new artificial intelligence system described in the journal Nature on September 4 takes many current artificial intelligence methods for cancer diagnosis.
Current artificial intelligence systems are often trained to perform specific tasks, such as detecting the presence of cancer or predicting the genetic characteristics of tumors, and they are often only suitable for a small number of cancer types. In contrast, the new model can perform a wide range of tasks and has been tested on 19 cancer types, giving it the same flexibility as large language models such as ChatGPT.
Although other basic artificial intelligence models for medical diagnosis based on pathological images have recently emerged, this is believed to be the first model to predict patient outcomes and be validated in multiple international patient populations.
“Our goal is to create a flexible, versatile ChatGPT-like AI platform that can perform a wide range of cancer assessment tasks,” said Kun-Hsing Yu, senior author of the study and assistant professor of biomedical informatics at Harvard Medical School’s Blavatnik Institute.
“Our model has proven very useful in multiple tasks related to cancer detection, prognosis and treatment response in a wide range of cancers.”
The artificial intelligence model detects cancer cells by reading digital slices of tumor tissue and predicts the molecular characteristics of the tumor based on the cellular characteristics seen on the image, with higher accuracy than most current artificial intelligence systems.
It can predict survival in patients with multiple cancer types and accurately identify the characteristics of the surrounding tissue (also known as the tumor microenvironment) that are correlated with patient response to standard treatments, including surgery, chemotherapy, radiotherapy and immunotherapy.
Finally, the team said the tool appears to be able to generate new insights-it identifies previously unknown specific tumor characteristics related to patient survival.
The research team said the findings provide further evidence that AI-driven methods can improve clinicians ‘ability to effectively and accurately assess cancer, including identifying patients who may not respond well to standard cancer therapies.
“If further validated and widely deployed, our method and methods similar to ours can identify cancer patients early who may benefit from experimental treatment targeting certain molecular variants, an ability that is not universally available around the world,” Yu said.
Training and performance
The team’s latest work builds on Yu’s previous research on artificial intelligence systems for evaluating colon cancer and brain tumors. These early studies demonstrated the feasibility of the method in specific cancer types and specific tasks.
The new model, called CHIEF (Clinical Histopathology Imaging Evaluation Foundation), was trained on 15 million unlabeled images that were divided into sections of interest. The tool was then further trained on 60,000 complete tissue section images, including lungs, breast, prostate, colorectal, stomach, esophagus, kidney, brain, liver, thyroid, pancreas, cervix, uterus, ovaries, testis, skin, soft tissue, adrenal glands and bladder.
The model is trained to view specific parts of the image and the entire image, allowing it to correlate specific changes in an area with the overall context. Researchers say this approach allows CHIEF to interpret images more comprehensively by considering broader contexts rather than just focusing on specific areas.
After the training, the team tested the performance of CHIEF on more than 19,400 full slide images from 32 independent datasets from 24 hospitals and patient cohorts around the world.
Overall, CHIEF performs 36% better than other state-of-the-art artificial intelligence methods in tasks such as cancer cell detection, tumor origin identification, predicting patient outcomes, and identifying the presence of genes and DNA patterns associated with treatment response.
Thanks to its versatile training, CHIEF performs equally well whether tumor cells are obtained through biopsy or surgical resection. No matter what technology is used to digitize cancer cell samples, it is equally accurate. Researchers say this adaptability allows CHIEF to be used in different clinical settings and represents an important step beyond current models, which often perform well only when reading tissue obtained through specific techniques.
cancer detection
CHIEF achieves nearly 94% accuracy in cancer detection and is significantly superior to current AI methods on 15 datasets containing 11 cancer types. Across five biopsy datasets collected from independent cohorts, CHIEF achieved 96% accuracy in multiple cancer types, including esophageal cancer, gastric cancer, colon cancer, and prostate cancer.
When researchers tested CHIEF on previously unseen slides of surgically removed colon, lung, breast, endometrial and cervical tumors, the model achieved an accuracy rate of over 90%.
Predicting the molecular spectrum of tumors
The genetic makeup of tumors holds key clues that determine their future behavior and optimal treatment. To obtain this information, oncologists require DNA sequencing of tumor samples, but due to the cost and time required to send samples to specialized DNA sequencing laboratories, detailed genomic analysis of such cancer tissue is not routinely or uniformly performed worldwide. Even in resource-rich areas, the process can take weeks. Yu said artificial intelligence can fill this gap.
Researchers say rapid identification of cellular patterns on images that imply specific genomic aberrations could provide a rapid and cost-effective alternative to genomic sequencing.
By observing microscope slides to predict tumor genomic mutations, CHIEF’s performance is superior to current artificial intelligence methods. This new artificial intelligence approach has successfully identified relevant characteristics of several important genes associated with cancer growth and suppression, and predicted mutations in key genes related to how tumors respond to various standard therapies.
CHIEF also detected specific DNA patterns that correlated with the degree to which colon tumors responded to an immunotherapy called immune checkpoint blocking.
When viewing full-tissue images, CHIEF identified mutations in 54 commonly mutated cancer genes, with an overall accuracy rate of more than 70%, which is superior to the most advanced artificial intelligence method for genomic cancer prediction. For specific genes for specific cancer types, the accuracy is higher.
The team also tested CHIEF’s ability to predict mutations associated with the response to FDA-approved targeted therapies across 15 anatomical sites. CHIEF has achieved high accuracy in multiple cancer types, including 96% accuracy when detecting EZH2 gene mutations that are common in blood cancers of diffuse large B-cell lymphoma. The treatment effect of BRAF gene mutations in thyroid cancer reaches 89%, and the treatment effect of NTRK1 gene mutations in head and neck cancer reaches 91%.
Predicting patient survival
CHIEF successfully predicted patient survival based on tumor histopathological images obtained at the time of initial diagnosis. Across all cancer types and all patient groups studied, CHIEF distinguishes between long-term and short-term survival patients.
CHIEF is 8% higher than other models. Among patients with more advanced cancer, CHIEF performed 10% better than other AI models. In summary, CHIEF’s ability to predict high versus low mortality risks was tested and confirmed with patient samples from 17 different institutions.
Extracting new insights into tumor behavior
The model identifies image patterns that are related to tumor aggressiveness and patient survival. To visualize these areas of interest, CHIEF generated heat maps on the images. When human pathologists analyzed these AI-derived hotspots, they saw interesting signals reflecting the interactions between cancer cells and surrounding tissue.
One feature is that long-term survivors have a greater number of immune cells in their tumor areas than short-term survivors. Yu pointed out that the finding makes sense because the increase in immune cells may indicate that the immune system has been activated to attack tumors.
When looking at tumors in short-term survivors, CHIEF identified regions of interest characterized by abnormal size ratios between various cellular components, more atypical features on the nucleus, weak connections between cells, and less connective tissue surrounding tumors in this region.
There are also more dying cells around these tumors. For example, in breast tumors, CHIEF identifies the presence of necrosis (or cell death) within tissue as an area of interest.
On the other hand, breast cancers with higher survival rates are more likely to retain a cellular structure similar to healthy tissue. The research team noted that visual characteristics and areas of interest related to survival vary depending on the type of cancer.
subsequent step
The researchers said they plan to improve CHIEF’s performance and enhance its functionality by:
Additional training on tissue imaging of rare and non-cancer diseases
Including pre-cancerous tissue samples before the cells become completely cancerous
Exposure of models to more molecular data to enhance their ability to identify cancers with varying levels of aggression
Trained models can predict benefits and side effects of new cancer treatments in addition to standard treatments
For more content, you can click on the link below the video
Thank you for watching this video. If you like it, please subscribe and like it. thank
Original text:https://medicalxpress.com/news/2024-09-chatgpt-cancer-treatment-choice-survival.html
More information: Pathological basic models for cancer diagnosis and prognosis prediction, Nature (2024). DOI：10.1038/s41586-024-07894-z 。Journal information: Nature
Oil tubing: