Article

Ongoing Challenges in CT: How AI Provides Possible Solutions

There has been a dramatic increase in computed tomography (CT) scans over the last 40 years with more than 80 million scans conducted in 2016 compared to just three million in 1980.1,2 Continuous technological innovations have had a revolutionizing impact on diagnosis and treatment increasing the range of CT applications and clinical uses.2,3 For example, invasive riskier procedures, such as exploratory surgeries, have been replaced by noninvasive CT imaging studies that provide accurate diagnosis.2,3

How AI could reduce radiation exposure

Despite the tremendous advances, challenges remain. Along with the rise in CT use has come concerns about patients' increased exposure to radiation.3,4 Responding to this concern, radiologists, medical physicists, and CT system manufacturers have begun focusing on the development of a range of techniques to reduce the radiation dose from CT.3,4 However, no technique to-date has been effective enough to overcome the image quality degradation that results from increased noise and artifacts characteristic of reduced radiation CT acquisitions.3,4

Salary Guide Banner.jpg

Conventional denoising techniques, such as the nonlocal means denoising algorithm and total variation minimization, to remove, maintain or improve diagnostic image quality are complicated to use and not particularly effective because of the complex unevenly distributed patterns of noise in low-dose CT.3,5,6 However, artificial intelligence (AI) methods, such as deep learning algorithms, are demonstrating their ability to overcome this issue because they are dependent on training samples instead of a noise type.6

Deep learning techniques are considered state-of-the-art for classification of images and have been shown to decrease error rates from 25 to less than four percent.7 More recently, improved computing power has allowed for the development of a range of AI solutions such as deep convolutional neural networks (CNN) that increasingly demonstrate the ability to preserve and enhance diagnostic image accuracy and quality.4 Successful studies include improved detection and classification of lung nodules, mediastinal lymph nodes, as well as segmentation of brain tissue.7

Improving time to treatment

Intracranial hemorrhage (ICH) is a time-sensitive condition and major factor in approximately 2 million strokes worldwide.8 ICH creates significant patient and healthcare system burden, including mortality rates of 47 percent with most deaths occurring during the first 24 hours of onset and permanent cognitive damage in up to half of survivors.8

Acute ICH is diagnosed by CT of the head with the time to scan and interpretation playing a major role in patient outcomes.8 Interpretation times can be impacted by the priority assigned to the exam and in some instances can be subject to significant delays when not prioritized.8 For example, the variation of interpretation times between inpatient and outpatient scans could be minutes compared to hours.8

Because every minute counts in maximizing ICH patient outcomes, one Pennsylvania-based integrated health system implemented a deep convolutional neural network (CNN) as a quality improvement tool for radiology workflow optimization.8 It was designed specifically to reduce time to diagnosis and treatment for outpatients with critical ICH, a patient group often facing greatest delays.8 Using the algorithms to automatically evaluate CT scans of the head for ICH during a three month prospective study they were able to re-prioritize routine imaging studies as “stat” in real-time radiology workflows if a hemorrhage was detected.8 Clinicians found a reduction of time to diagnosis for new outpatient ICH cases of 96 percent with improved detection of more subtle ICH that may be overlooked by radiologists.8

During the study period, 94 of 347 “routine” CT studies were re-prioritized to “stat” in radiology worklists by the deep learning algorithm.8 Of those 94 outpatient CT scans, ICH was identified by the radiologist in 60.8 Additional results include the detection of five new ICH cases, including one outpatient with vague symptoms on anticoagulation, and a reduction in median time to diagnosis to 19 minutes from 512 minutes.8

CT studies of more than 31,000 unique adult patients were collected retrospectively from the picture archiving and communication (PACS) system of the integrated healthcare system to develop, train, and test the deep CNN.8 None of the more than 46,000 non-contrast head CT studies had been used in research before and none were taken from public datasets.8 Each imaging study included was required to have a completed clinician dictation report as well as at least 20 axial 2D slices.8 All studies were acquired from facilities across the health system using 17 scanners from four different CT system manufacturers between 2007 and 2017.8

Deep learning optimizes screening, predicts risk

While lung cancer screening using low-dose CT imaging is reported to decrease mortality by up to 43 percent and is included in U.S. guidelines, there remain several challenges.9 For example, dozens of images within a single CT scan must be reviewed in order to identify cancerous nodules or tissue that are inherently difficult to identify and classify.10 Radiologist variability and high false-positive and false-negative rates, along with fatigue and distraction are also real issues to consider.11,9 Manually, lung cancer CT scan analysis is a time-consuming, inefficient, and arduous task that requires radiologists to spend hours detecting the presence of small spherical nodules in CT images that must undergo further evaluation to determine malignancy.12

Regardless, early accurate detection and evaluation of lung nodules in CT screening remains the best strategy for diagnosing lung cancers and preventing death.12 Given this, one team of researchers set out to build a deep learning algorithm that uses a patient’s current and prior CT studies to predict their risk of lung cancer.9 Their goal is to use deep learning and AI to boost a patient’s chance of survival.9

A model using nearly 46,000 de-identified chest CT screening studies was consolidated and compared to the performance of six board-certified radiologists.10 When using a single CT scan for diagnosis, the deep learning model achieved a state-of-the-art success rate of more than 94 percent area under the curve (AUC) and performed the same or better than human radiologists.9,10 In cases where no previous CT imaging was available, the deep learning model outperformed all of the six radiologists including absolute reductions of 11 percent of false positives and 5 percent of false negatives.9,10

The deep learning algorithm can determine overall lung malignancy, detect subtle malignant tissue, and incorporate data from prior scans since tissue growth rate may indicate malignancy.10

Despite the reality that most patients go unscreened, researchers believe results produced by their deep learning method highlights an opportunity to optimize the accuracy and consistency of the lung cancer screening process around the world using computer assistance and automation.9

While AI has already shown promise in addressing a number of key clinical challenges in CT and radiology in general, further progress is likely to occur in the coming years as the technology and algorithms become more sophisticated.

REFERENCES:

  1. IMV 2018 CT Market Outlook https://imvinfo.com/product/2018-ct-market-outlook-report/Accessed 5/20/2019
  2. Radiation risk from medical imaging. Harvard Health Publishing Harvard Medical School Harvard Women's Health Watch https://www.health.harvard.edu/cancer/radiation-risk-from-medical-imagingAccessed 5/20/2019
  3. High quality imaging from sparsely sampled computed tomography data with deep learning and wavelet transform in various domains. American Association of Physicists Onlinehttps://aapm.onlinelibrary.wiley.com/doi/pdf/10.1002/mp.13258Accessed 5/20/2019
  4. Basics of iterative reconstruction methods in computed tomography: A vendor-independent overview. European Journal of Radiology https://www.ejradiology.com/article/S0720-048X(18)30374-7/fulltextAccessed 5/20/2019
  5. Deep Convolutional Approach for Low-Dose CT Image Noise Reduction. IEEE Xplore https://ieeexplore.ieee.org/document/8430255Accessed 5/20/2019
  6. Low-dose CT via deep neural network. Biomedical Optics Express https://arxiv.org/ftp/arxiv/papers/1609/1609.08508.pdfAccessed 5/20/2019
  7. Deep Learning at Chest Radiography: Automated Classification of Pulmonary Tuberculosis by Using Convolutional Neural Networks. RSNA Radiology Journal https://pubs.rsna. org/doi/full/10.1148/radiol.2017162326Accessed 5/20/2019
  8. Advanced machine learning in action: identification of intracranial hemorrhage on computed tomography scans of the head with clinical workflow integration. Nature https://www.nature.com/articles/s41746-017-0015-zAccessed 5/20/2019
  9. End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography. Nature https://www.nature.com/articles/s41591-019-0447-xAccessed 5/20/2019
  10. Google Develops Deep Learning Tool to Enhance Lung Cancer Detection. Health IT Analytics https://healthitanalytics.com/news/google-develops-deep-learning-tool-to-enhance-lung-cancer-detectionAccessed 5/20/2019
  11. An Appraisal of Lung Nodules Automatic Classification Algorithms for CT Images. Sensors (Basel) https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6338921/Accessed 5/20/2019
  12. Classification of lung nodules in CT scans using three-dimensional deep convolutional neural networks with a checkpoint ensemble method. BMC Medical Imaging https://bmcmedimaging.biomedcentral.com/articles/10.1186/s12880-018-0286-0Accessed 5/20/2019