Harnessing AI to Revolutionize Cancer Diagnosis
Artificial intelligence (AI) is changing the way we interpret information. It is enabling machines to better analyze and characterize even the most complex data, thereby facilitating superior experiences, services and decision making. In particular, the harnessing of AI to revolutionize healthcare and medical applications has seen significant progress in recent years.
The advent of AI in cancer screening and diagnosis shows tremendous promise for facilitating and supporting the early detection of cancer, critical for more effective treatment. While the field of AI for cancer diagnostic tools is still gaining momentum, it is clear that such technological innovation will revolutionize the world of cancer screening, detection and diagnosis.
In this article, we will review four AI-based decision-support tools for cancer screening, detection, and diagnosis, of which three are still in clinical testing or experimental stages and one is already deployed in a live clinical setting.
Artificial intelligence improves lung cancer detection in CT scans
Low-dose computed tomography (CT) scans have, in recent years, been proven quite effective in earlier detection of lung cancer, reducing mortality from lung cancer by up to 20 percent1.
With the desire to make CT scans even more effective, three global high-tech companies, Intel, Alibaba Cloud and LinkDoc, announced an innovation competition, specifically designed to deal with the rising challenge of lung cancer in China.
The goal was to develop algorithms and apply AI to CT technology to enhance its contribution to the detection of small pulmonary nodules. Participants, mainly medical technology researchers and developers, built and trained AI models that identify lung cancer nodules at high accuracy. To train the models, participants used a database of around 3,000 clinical records, CT images and imagery evaluations.
The winning team, from Peking University, implemented a 3D convolutional neural network (CNN) model - an artificial neural network that mimics neurological processes related to the intake and processing of visual information. The 3D CNN model was trained with data in varying resolutions and the conclusion was that higher resolution data achieves better detection performance. This means that high resolution data must be used in order to detect very small nodules that could be missed by humans.
This impressive work still requires a lot of research and experimental work to mature, but there is no doubt that the entire field of cancer diagnosis profited, as data collected during the competition supported the hypothesis that AI improves the accuracy of detecting smaller pulmonary nodules, which can be challenging for human radiologists to identify.
AI increases accuracy of CADe for mammography
Computer Assisted Detection (CADe) is an indispensable tool for healthcare analytics. When paired with mammography, CADe assists radiologists in identifying abnormal breast tissue areas.
The analysis of mammography images is time consuming and can create physical and mental strain on any physician. Researchers from Hungary’s Eötvös Loránd University found that integrating deep learning into existing CADe programs, in order to better identify meaningful visual patterns and representations associated with breast pathology, can increase accuracy and significantly reduce error rates, as well as costs, compared to other technologies. The system needs to be further trained, using additional, larger data-sets, with the goal of eventually bringing it to market, making it an integral part of routine life-saving medical care.
AI-based cancer diagnostics system for prostate cancer; the first ever to be deployed clinically in a pathology lab
Seeking to empower pathologists to make faster, more accurate cancer diagnoses, Ibex Medical Analytics has developed computer software that identifies various cell and tissue types from within whole slide images of prostate core needle biopsies (PCNBs), including the grading of cancerous glands and other clinically significant features. Currently implemented for prostate cancer diagnosis, the Ibex algorithm utilizes state-of-the-art AI and Machine Learning techniques, and was trained on many thousands of image samples, taken from hundreds of PCNBs from multiple institutes.
Following a successful pilot period, in which the Ibex Second Read™ (SR) system identified isolated major errors in retrospective PCNBs, previously diagnosed by human pathologists as benign, the system was deployed within the pathology institute of Maccabi Healthcare Services, one of Israel’s largest healthcare providers.
Soon after deployment, the system identified a suspicious PCNB, reported as benign by a pathologist just hours earlier. It was subsequently re-examined and confirmed to be a low-grade prostate cancer (adenocarcinoma), creating a significant and immediate impact on prostate cancer diagnosis and patient care. The Ibex SR system is the first ever AI-based cancer diagnostic tool to be up and running in a live clinical environment.
AI algorithms train computers to screen for skin cancer
Skin cancer is the most common type of cancer. Hence, the ability to use technology to ease the significant workload on dermatologists and pathologists without compromising and even improving the ability to accurately screen for skin cancer, can make a real difference in improving early detection and decreasing mortality rates. Also, such technology could reduce the need to pay unnecessary visits to the dermatologist, lessening the existing burden on the healthcare system.
A team of researchers from Germany, the United States and France launched an experiment involving a deep learning CNN and 58 dermatologists. The researchers exposed the CNN to over 100,000 images and taught it to distinguish malignant skin lesions from benign ones. They found that the CNN accurately identified 95 percent of skin cancers, compared to only 86.6 percent correctly identified by human dermatologists
Roughly half the dermatologists were classified as field experts, with over five years of experience under their belts. 19% of dermatologists had between two and five years of experience, and 29% were beginners, with less than two years of experience practicing dermatology. Yet, most were outperformed by the CNN, indicating that AI is expected to become a key tool for faster and easier diagnosis of skin cancer, enabling malignancies to be surgically removed before they spread.
The healthcare industry is constantly on the cusp of new and exciting innovation. In this article we reviewed four state-of-the-art AI-based decision support tools for cancer screening, detection and diagnosis. The journey towards using AI in mainstream decision-support tools is well underway, and we expect to see many additional AI-based applications. Within the next 3-5 years, we believe that AI will revolutionize cancer diagnosis, generating objective and reproducible results and decreasing diagnostic error rates by deploying a clinical safety net, and most importantly - significantly impacting patient care for the better.
- Aberle D.R., Adams A.M., Berg C.D., et al. “Reduced Lung-Cancer Mortality With Low-Dose Computed Tomographic Screening.” New England Journal of Medicine, Aug. 4, 2011. DOI: 10.1056/NEJMoa11028732