Contents
Cancer is the second most common cause of death in our country. Machine learning improves the accuracy of detecting and predicting cancer susceptibility. Trends figured out how it works
The International Agency for Research on Cancer (IARC) estimates that one in five people will develop cancer during their lifetime. In 2019, over 640 people with cancer were diagnosed in our country alone. Experts attribute the increase in mortality from cancer to the general aging of the population. The chances of survival in cancer patients depend on the stage at which the cancer is detected: the sooner the better.
Artificial Intelligence: An Evolution in Cancer Diagnosis
Artificial neural networks and automatic data analysis have been used to detect and diagnose cancer since the mid-1980s. Today, the most innovative methods are based on machine learning.
What is the difference between machine learning and deep learning
Machine learning (Machine Learning, ML) is an artificial intelligence method that is used to analyze data. The analysis algorithm builds a single mathematical model using the neural network “training” process, which is similar to teaching children.
The child recognizes letters, learns to put them into syllables, and syllables into sentences. The learning process of a neural network is also based on the initial data elements on which the network learns to solve some problem.
Deep learning (Deep Learning, DL) is a type of machine learning based on neural networks. The main difference between deep learning and machine learning is based on how data is presented to the system. Machine learning algorithms almost always work on structured datasets, and deep learning networks rely on their own layers of ANNs (Artificial Neural Networks) to structure the data themselves.
The error of both machine learning and deep learning models depends on the quality of the data.
Opinions are divided on the maturity and success of machine learning in medicine for analyzing CT images, MRI, mammograms and so on for cancer detection. The total number of projects and positive scientific publications on cancer ML diagnostics is growing, but, for example, IBM, as one of the first developers who tested IBM Watson Oncology in more than 50 clinics, reduced the unit’s staff in 2020, and in 2021 announced about the intention to sell it.
Despite the wave of skepticism, the number of developments and products for the detection of malignant neoplasms is increasing, including those recommended by the FDA (US Food and Drug Administration). For example, Philips Healthcare and SIEMENS Healthineers are considered leaders in lung cancer screening. Google AI Healthcare and IBM Watson Oncology are also popular, though not recommended by the FDA. There is a lot of competition in the ML solutions market from startups and open source projects. To date, the FDA has approved medical use over 80 ML solutions.
Examples of using ML solutions
- Mia algorithm for breast screening of the English project Kheiron Medical Technologies has proven itself well as a second opinion. The first results of more than 40 mammograms showed that if Mia were used as a second radiographer, the overall double-read repetition rate (percentage of women who are called for further examination) would be 4-5%, and the cancer detection rate would be 8,4 people per 1 thousand patients.
- Gleason scale is a popular prostate cancer scoring system developed back in 1966. Physicians have historically evaluated tissue biopsies visually. And only in 2019, Martin Stump (from AI and Data Science) and Craig Mermel (from Google AI Healthcare) developed a deep learning system for the Gleason scale. Diagnostic accuracy was 0,7 (on a scale of 0,5 “random” to 1 “100% correct”). This is an excellent result compared to the diagnosis made by 29 doctors, whose accuracy is 0,6.
Evgeny Nikitin, Head of the Artificial Intelligence Department, Medical Screening Systems LLC (Celsus):
“Good clinical performance of ML systems does not guarantee the benefits of their implementation. In each specific task, you need to optimize your own metric – for example, the time spent on describing one study, the financial costs of patients and hospitals, the percentage of detected neoplasms that require emergency medical care. So far, there are very few full-fledged success stories of the implementation of ML systems, but the first results give hope: for example, the results of our experimental tests show that the use of ML systems speeds up the description of a mammography examination by 30-40%. And clinical trials of our Celsus system at the Tambov Regional Oncological Clinical Oncology Center showed an increase in the detection of oncology at an early stage by 10%.
How ML Cancer Diagnosis Works
A large number of images (obtained using CT, MRI, mammography or histopathology) are collected in datasets – datasets that have become used to train machine learning algorithms.
The image is loaded into a system that prioritizes a list of studies from the highest probability of pathology to the lowest. So the doctor will first look at the pictures of patients in whom the system predicted a neoplasm. Or the specialist looks at the image, where the AI highlighted the area of pathology with a marker, and makes his remarks in the description of the image made by the AI.
The process of creating an ML model by developers consists of many stages – data collection and labeling, image preprocessing (for example, organ segmentation and removal of unnecessary parts of images), neural network training, and results calibration for a specific application scenario.
One of the most difficult stages is segmentation (feature extraction) for classification into a malignant or benign tumor. Different imaging and training processes are used to recognize and segment each type of cancer, which are different for diagnosing, for example, skin cancer and leukemia.
How ML Cancer Diagnosis Is Tested
Since there are many algorithms and datasets, there is a problem of comparing ML solutions developed for the same clinical application.
What metrics are used to compare ML solutions
Hundreds of different metrics can be used to evaluate machine learning systems. The easiest to understand are metrics based on the breakdown of AI decision predictions into four categories:
Machine learning is being actively tested to detect brain, breast, lung, skin, blood and liver cancers. In our country, in 2020, the Center for Diagnostics and Telemedicine launched a prospective experiment to test AI solutions. It became the largest scientific study in the world on this topic. Algorithms for detecting breast cancer and lung cancer were included in the Moscow testing program in the summer of 2020 after services for diagnosing COVID-19.
Natalya Ledikhova, Deputy Director for Medical Affairs of the Scientific and Practical Clinical Center for Diagnostics and Telemedicine Technologies of the Moscow Department of Health, told Trends that, according to the results of AI testing to detect breast cancer, the time to describe the study is reduced by 15–50%. In addition, algorithms process medical images in real time and prioritize them: the doctor receives the first examinations that are more likely to detect oncological changes. ML-services focus the doctor’s attention on the most insignificant signs of pathology – this can help with early cancer diagnosis, especially with large streams of routine studies, when a specialist can “blur” his eyes. The accuracy and sensitivity of artificial intelligence is constantly growing.
In 2020, the developments of Philips (LDCT of the chest, lung cancer), Botkin AI (CT of the chest, lung cancer), Celsus and Lunit (mammography, breast cancer) were used to diagnose oncological diseases in our country. In the future, oncological testing may be expanded to include services for the detection of malignant brain tumors.
Before connecting services to the main circuit of the Unified Radiological Information Service (ERIS), which is used by Moscow radiologists, ML solutions are internally validated on independent data arrays collected and structured by the Center for Diagnostics and Telemedicine (reference sets of anonymized studies previously marked up by radiologists) in a test circuit service. At the moment, more than 100 datasets of radiation diagnostics studies have been prepared. This is done to make sure that the AI solution, trained and tested by the developer on a specific data set, will be able to work well in the new conditions.
AI Mistakes: Who to Blame for Misdiagnosis
Every year 70 people die in our country due to mistakes and unprofessionalism of doctors. In America, this is one in three deaths. In our country, there is no established practice of blaming a doctor for an incorrect diagnosis. The use of AI raises even more serious questions about negligence, as not the best medicine can become even worse. If AI harms a patient due to a misdiagnosis, then who will be held accountable – the developer of the algorithm, the clinic, the doctor operating the algorithm, or the government agency that authorized the use of the algorithm?
So far, ML-diagnostics is an auxiliary tool for making clinical decisions, and not a replacement for medical diagnostics, so the physician is responsible for possible errors.
At the moment, it is difficult to understand, for example, why AI issued such a recommendation, due to the opacity of the work of machine learning algorithms. Now they work as “black boxes” – many algorithms are confidential information of developers, as well as patient data sets for training algorithms.
Barriers to mass adoption of ML solutions
In addition to the difficulty of comparing and choosing machine learning tools, there are a number of other problems:
- Technological barrier
The main problem is the lack of large datasets with good (clean and complete) data. There are also no reference datasets on which it would be possible to compare the accuracy of the algorithms of different vendors. As long as there are no such datasets, the solutions will be narrow (to identify an anomaly in the image – yes; to generalize different analyzes and make a diagnosis – no, only a doctor can do this for now). Public policy can promote the implementation of such solutions through the formation of datasets and the creation of pilot zones.
- Complexity of integration
Experts predict a complex integration of ML diagnostics into the already established practice of collecting and processing images in hospitals. Our country now has a large number of different systems, and developers of AI solutions need to spend a significant part of their resources on integration activities.
- Russian regional specifics
In addition to the complexity of integration, many regions of our country have not yet expressed an active desire to conduct pilot tests and implement AI solutions. In Moscow, the organizers of the experiment were prepared for the fact that all medical ML products are still far from ideal and often require significant refinement, additional training and calibration. In the regions, there is much less scientific and experimental interest in ML products, they are waiting for finished solutions. In addition, skepticism and distrust of doctors in AI in the regions is much higher. That is, only developers with clinically proven systems will be able to work successfully in the regions.
- Registration certificate of a medical device
Sale and operation of ML-solutions in the field of healthcare can be carried out only after obtaining a registration certificate for a medical device issued by Roszdravnadzor.
Evgeny Nikitin:
“The specifics of medical products based on artificial intelligence are associated with a long development cycle before reaching operational payback and a high level of costs for preparing and labeling medical data. As a rule, now companies raise only in the first investment round from ₽50 million to ₽180 million.”
Cancer Prediction
Despite the growing reliance on protein biomarkers for cancer diagnosis and the bias in favor of detecting prostate and breast cancer, machine learning provides a significant (15–25%) improvement in the accuracy of predicting cancer predisposition, recurrence, and mortality.
In 2019, a team of IBM researchers published a study on a new AI model that can predict the development of malignant breast cancer in patients over the next year. This is the first algorithm that takes into account not only images, but also the patient’s health history. The model correctly predicted the development of breast cancer in 87% of cases, and correctly interpreted 77% of non-cancerous lesions.
Cancer Prediction Trends:
- predicting predisposition to cancer (that is, risk assessment) – the likelihood of developing any type of cancer before the onset of the disease;
- prognosis of cancer recurrence – the likelihood of recurrence of cancer after a visible cure for the disease;
- survival prognosis – to predict the outcome (life expectancy, survival, progression, tumor drug sensitivity) after diagnosis.
Personalized Cancer Treatment
Cancer patients need complex individual complex treatment. AI-based solutions can transform not only diagnostics, but also the field of patient care:
- Genomic characterization of tumors
Artificial intelligence techniques are being used to identify gene mutations from tumor images instead of using traditional genome sequencing. For example, researchers at New York University used deep learning to analyze images of lung tumors and found that DL can not only accurately distinguish between the two most common subtypes of lung cancer, adenocarcinoma and squamous cell carcinoma, but also predict mutated genes from images.
- AI-assisted personalized drug therapy
Researchers at Aalto University, the University of Helsinki and the University of Turku in Finland have developed a machine learning model that accurately predicts how combinations of different cancer drugs kill different types of cancer cells. Complicated drug therapy often increases the effectiveness of treatment and can reduce harmful side effects if the dosage of individual drugs is reduced. With the help of machine learning, it is possible to find the best combinations for the selective destruction of cancer cells with a certain genetic or functional composition.
- AI-assisted customization of radiation therapy doses
Radiation therapy is very difficult to tolerate by the body and it is difficult to predict how patients will react. In 2019, AI was successfully used to plan the focus of radiotherapy and accurately predict its side effects in patients with head and neck cancer. The new technology will allow doctors to better plan radiation therapy for each individual.
- Predicting Immunotherapy Outcomes
According to the study, an AI algorithm can successfully find previously unseen changes in CT scan image patterns and determine how patients with lung cancer will benefit from immunotherapy. Immunotherapy is an expensive treatment, but currently only about 20% of cancer patients actually benefit from it.