preface
According to the latest World Health Organization survey report, in 2022, there were an estimated 20 million new cancer cases and 9.7 million deaths. The estimated number of people surviving within 5 years of cancer diagnosis is 53.5 million. About 1 in 5 people will develop cancer in their lifetime, and about 1 in 9 men and 1 in 12 women will die from cancer.
On February 2, 2024, the World Health Organization's International Agency for Research on Cancer (IARC) recently released the Global cancer burden growing, amidst mounting need for services, which projected more than 35 million new cancer cases in 2050, a 77% increase from the estimated 20 million cases in 2022. This once again underscores the growing global cancer burden, which deserves worldwide attention.
Histopathological image assessment is an effective method for diagnosing cancer. Recently, a research team from Harvard Medical School and its collaborators proposed the Fundamentals of Clinical Histopathological Imaging Assessment (CHIEF) model for extracting pathological imaging features for systematic cancer evaluation.
On 15 datasets containing 11 cancer types, CHIEF achieved nearly 94% accuracy in cancer detection, significantly outperforming current AI methods. In 5 biopsy datasets collected from independent cohorts, CHIEF achieved 96% accuracy across multiple cancer types, including esophageal, gastric, colon, and prostate cancers. When researchers tested CHIEF on previously unseen sections of surgically removed tumors from the colon, lung, breast, endometrium, and cervix, the model was more than 90% accurate.
相关研究论文以“A pathology foundation model for cancer diagnosis and prognosis prediction”为题，已发表在权威科学期刊 Nature 上。
"Our goal is to create a flexible, versatile, ChatGPT-like artificial intelligence (AI) platform that can perform a wide range of cancer assessment tasks, and our model is very useful in multiple tasks related to cancer detection, prognosis, and treatment response across multiple cancers," said Kun-Hsing Yu, co-corresponding author of the study and assistant professor at Harvard Medical School. ”
The researchers noted that if the method is further validated and widely promoted in the future, it will be possible to identify patients with early-stage cancer. These patients may benefit from experimental treatments targeting specific molecular variants, which will help close the gap in the development and application of such treatments globally.
Up to 94% accuracy in detecting cancer
CHIEF is a general-purpose machine learning framework for weakly supervised histopathological image analysis. CHIEF extracts pathological imaging findings useful for cancer classification, tumor origin prediction, genomic prediction, and prognostic analysis. The research team pre-trained CHIEF in a weakly supervised manner using 60,530 full-slice images representing 19 anatomical sites.
During the pre-training process, they cropped the whole slice image into non-overlapping image tiles, and used the Contrastive Language-Image Pre-Training (CLIP) embedding method to encode the anatomical site information of each whole slice to obtain the feature vectors of each anatomical site. They combined text and image embeddings to represent heterogeneous pathological information from the training data. Then, the type of cancer is directly inferred using the pathological imaging features extracted by CHIEF. In genomics prediction and prognostic prediction tasks, the CHIEF feature serves as the basis for fine-tuning the model for each specific task.
图 | CHIEF 模型概述。
CHIEF outperforms state-of-the-art deep learning methods by up to 36.1% on these tasks. On average, CHIEF outperformed traditional methods by 9%.
Figure | CHIEF significantly outperforms state-of-the-art methods in cancer classification, genomic identification, and survival prediction tasks.
The CHIEF model has demonstrated strong versatility and generalization in pathological image analysis, and can be applied to a variety of pathology evaluation tasks in a variety of cancer types, including cancer detection, tumor origin prediction, genomic feature prediction, and survival prediction.
CHIEF achieved a macroscopic mean area under the receiver operating characteristic curve (AUROC) of 0.9397 on 15 datasets representing 11 cancer types, which is 10% or more higher than all three existing deep learning methods. In all five biopsy datasets collected from independent cohorts, CHIEF achieved AUROCs greater than 0.96 in several cancer types, including esophagus, stomach, colon, and prostate. AUROCs of CHIEF greater than 0.90 when independently validated using seven sets of surgically resected sections covering five cancer types (i.e., colon, breast, endometrium, lung, and cervix). These results demonstrate the ability of CHIEF to generalize in diverse cancer tissues and samples from diverse sources internationally.
Figure | CHIEF outperforms state-of-the-art deep learning methods.
AI is helping humanity beat cancer
In the field of healthcare, AI is gradually showing its unique value; Especially in the early screening and detection of cancer, the application of AI technology is increasingly becoming a key force in overcoming this problem. The rapid development of this field is constantly driven by the continuous emergence of research results.
In June, a team of researchers from Imperial College London and the University of Cambridge jointly trained a new type of AI model, EMethylNET. By analyzing DNA methylation patterns, the model was able to accurately identify 13 different types of cancer, including breast, liver, lung, and prostate cancers, in non-cancerous tissues, with a detection accuracy of up to 98.2%, providing strong technical support for early cancer detection.
In July, a team of researchers at Harvard Medical School worked with partners to develop PathChat, a vision-language universal AI assistant for human pathology. The system is nearly 90% accurate in correctly identifying diseases when processing biopsy sections, outperforming the current market's general-purpose AI models such as GPT-4V as well as specialized medical models. The research paper has been published in the scientific journal Nature.
In addition, a research team is committed to using AI technology to manipulate cell fate and achieve a breakthrough in transforming cancer cells into immune cells. In August, scholars at the Keck School of Medicine at the University of Southern California (USC) conducted an innovative study with funding from the National Institutes of Health (NIH) in United States. They used AI to identify and reprogram the genes of glioblastoma cells, turning them into dendritic cells with anti-cancer capabilities, effectively targeting and destroying surrounding cancer cells. In a mouse model of glioblastoma, this approach significantly improves the chances of survival by 75%, providing a new perspective on cancer treatment.
At the same time, considering the problem of drug tolerance, some research teams have shifted their focus to metastatic cancer, using AI technology to develop personalized cancer treatment strategies. In June, scientists from research institutes such as the University of Oxford, through interdisciplinary joint research, introduced a novel framework that enables the application of deep reinforcement learning to develop adaptive treatment plans for prostate cancer patients. The results of the study show that this novel adaptive approach can significantly extend the time to recurrence-free in patients by up to two times compared to reliance on maximally tolerated doses or non-personalized intermittent therapy, opening up a new avenue for personalized cancer treatment.
Perhaps in the not-too-distant future, with the help of AI, humans will no longer talk about "cancer".
Source: Academic Headlines
Written by | Ma Xuewei
Editor | Academic Jun