February 4 th: World Cancer Day. Globally, nearly one in six deaths worldwide is caused by cancer, and nearly 70 percent of cancer deaths occur in low- and middle-income countries. Cancer caused 8.8 million deaths in 2015, with the most common types of cancer being lung, liver, colorectal, stomach and breast cancer. In China, the burden of cancer is also increasing year by year, with an average of more than 10,000 people diagnosed with cancer every day in 2015 and 7.5 people diagnosed with cancer every minute.
In breast cancer, for example, mammography is the “golden standard” for screening for breast cancer. Even so, reading X-rays is a difficult task for medical experts, with often examples of false positives (misdiagnosis) and false negatives (missed). This not only causes a heavy workload for doctors, but also delays the treatment of patients and puts them under unnecessary stress.
Now, with “AI-assisted medicine” and the topic of “AI-driven medicine” being frequently raised, it also means that technology is taking on an increasingly important role in the medical and health industries, such as assisting doctors to break through the ceiling of previous medical levels and replace doctors with repetitive routines. Or improve the imbalance in the regional medical allocation and so on.
Drugs for the disease
“Over the past few years, the Google team has applied AI to healthcare, from predicting patient diseases through the analysis and study of electronic medical records to assisting the detection of lung cancer, and while we are still in the early stages of technology development, the results are promising. Google CFO Ruth Porat says.
Last May, Nature Medicine published new developments in Google’s lung cancer testing, which predicts lung cancer based on low-dose computer tomography images.
Radiologists can’t review 3D scans like computers, and they need to review hundreds of 2D images to find problems. Google has created machine learning models that analyze high-pass 3D images, generate overall tumor predictions, and identify subtle, malignant tissues. By entering the patient’s previous CT image, the model analyzes and evaluates the growth rate of the suspect pulmonary nodules.
The Google team trained with 45,856 unidentified CT images and compared the results with six certified radiologists. Without the assist of a radiologist, the Google model detected a 5% reduction in false negatives and an 11% reduction in false positives.
Google’s efforts to diagnose breast cancer began earlier. In general, the way breast cancer cells spread is usually transferred to nearby lymph nodes, which affect treatment decisions for radiation therapy, chemotherapy, and surgery to remove additional lymph nodes. At least half a million people worldwide have died of breast cancer in the past, 90 percent of them metastatic tumors.
Detection of cancer stoading from the primary site to nearby lymph nodes is an important and difficult step in pathological examination. Most cancers involve the detection of lymph node metastasis, which is one of the basis for the widely used TNM cancer phase.
Google took the tool LYNA (LYmph Node Assistant) to the 2016 ISBI Camelyon Challenge, a competition focused on the classification and location of pathological slices of metastasis of breast cancer in the lymph nodes.
In 2018, Google published two separate papers on progress in breast cancer. In the first paper, Google applied the LYNA algorithm to identify pathological slices of camelyon Challenge and independent data sets provided by the paper’s co-authors. LYNA has been shown to be consistent in image variability and histological artifacts, and achieves similar performance on both data sets without additional research and development.
Left: Carrier slots containing lymph nodes have multiple histological artifacts Right: LYNA identifies the tumor area in the center (red) and correctly classifies the surrounding pseudo-filled area as a non-tumor area (blue)
In two data sets, LYNA was able to distinguish between slide slides with metastatic and non-metastatic cancers at 99% correct ness. In addition, LYNA can determine the location of cancers and suspected cancers in each slide, some of which cannot be detected by pathologists because they are too small. So the Google team speculated that an important use of LYNA was to highlight these “suspicious” areas and assist pathologists in making a final diagnosis.
In the second paper, six certified pathologists examined the lymph nodes of metastatic breast cancer with the assistance of LYNA and without LYNA assistance. Thanks to LYNA, pathologists have halved the average diagnosis time, checking each slide for only one minute, and the pathologist subjectively considers the diagnosis “easier” with lyNA’s help. In terms of diagnostic accuracy, with the help of LYNA, pathologists halved the missed rate of micro-metastasis of the lymph nodes.
Left: Magnified image of carrier slivers with micrometustatic lymph nodes Right: same view, with LYNA assisted to indicate the location of the tumor in blue
These advances may sound exciting, but more in the scientific trial phase, with a limited database, simulated diagnostic workflow, a separate examination of the pathopathic slides of individual lymph nodes in each patient rather than multiple lymph node pathologic slides commonly found in actual clinical cases, all of which allow LYNA Algorithms are a long way from true clinical practice.
Happily, at the start of 2020, Google is bringing good news to cancer diagnosis. On January 1, Google Health teamed up with DeepMind to publish a breast cancer artificial intelligence detection system in the Journal nature. The model was trained and adapted on a representative data set of more than 76,000 British women and more than 15,000 American women with unrecognized mammograms. The assessment was then made on a separate, unidentified data set (including more than 25,000 British women and more than 3,000 American women). The results showed that compared to radiologists, the AI model had a 5.7% lower false positive (United States) and 1.2% (United Kingdom), and a false negative was 9.4% (United States) and 2.7% (United Kingdom). In another study, the system outperformed six radiologists.
Detection performance of artificial intelligence detection system for breast cancer
Whether it’s the LYNA algorithm or this breast cancer detection system, Google’s research shows that the best outcomes at this stage come from a joint effort of professional humans and technology. For example, the UK’s breast screening process involves two doctors reading the tablets together, and in this case, the researchers asked the system and human experts to make the first decision at the same time, agreeing that they could not read the second time, and that the disagreement would start the second reading. The researchers found that the aI system maintained its non-performing performance, reducing the second reader’s workload by 88% compared to traditional “double reading.”
According to Google, the future of this artificial intelligence detection system has profound implications for clinical medicine. To verify that the model can be extended to other populations and screening programmes. The Google team simply retrained the system with UK data and evaluated it in US data. In this experiment, the AI model performed better than human experts, with a 3.5% reduction in false positives and an 8.1% reduction in false negatives. Although the gap has narrowed slightly, tests have shown that in future clinical deployments, the system may provide a strong foundation, improve the accuracy and efficiency of cancer screening, reduce patient wait times and stress, and improve the performance of the model by fine-tuning local data. But to achieve this goal, researchers still need ongoing research, prospective clinical trials, and regulatory approval.