Article

Could AI Improve Radiation Oncology Contouring Accuracy?

Could AI Improve Radiation Oncology Contouring Accuracy?

The advent of what some call the "fourth industrial revolution"—artificial intelligence (AI)—has brought about a revolution and transformed the more established industries, such as manufacturing and transportation.1,2 While the ongoing debate about the intrusion of machines into our lives and the replacement of human jobs is still at hand, developments in AI and robotics have far-surpassed expectations, and, in fact, even streamlined certain areas in medicine.3

For instance, in the radiology space, various applications of AI have been a game-changer for electronic medical records, imaging, or the management of outcomes or pathology data. Technological advancements, by way of adoption of intricate disease models and strategies, have vastly improved disease prediction. These newer machine-learning methodologies have largely contributed to the improved quality of cancer care and treatment.1 

Deep-learning algorithms for contouring of tumors

In 2012, Krizhevsky et al presented their breakthrough study that used a convolutional neural network (CNN) model, AlexNet, to reduce the error rate for object recognition. In time, this model, which showed record-breaking results, became important in shaping the future of radiotherapy and planning. Following this, a related, also successful, research especially focused on the use of CNN to inform automatic multi-organ segmentation in patients with head and neck cancer.4,5,6 

Picking up from here, earlier this year, a paper published in Radiology determined an artificial intelligence algorithm that could quickly and accurately provide contouring of primary gross tumors on MRI scans of patients with nasopharyngeal carcinoma. The tool could better help treat patients with this rare form of cancer. Precise and advanced AI-contouring such as this may even have a positive effect on patient mortality in the future.7

“Contouring accuracy is clinically important, as sub-optimal tumor coverage and poor-quality radiation therapy plans are major factors for disease relapse and inferior survival,” the first authors of this study explained.8

Tumor contouring, a process that determines cancerous volume periphery for potential surgery or radiation therapy, has been around for a while and was performed manually by doctors or specialists—until automation came along. Experts believe where manual contouring could be time-consuming and subjective, auto-contouring is faster and more standardized. In retrospect, it was a decade ago that Diane Heaton, a radiation oncologist at the Oklahoma CyberKnife cancer center, said, “I believe it [auto-contouring] is quicker and more accurate in many instances. This is an area in medicine that is literally exploding. I think we’re going to see computerized mapping, computerized identification structures, and [it will probably be used in] both the diagnostic and therapeutic realms.”9

In a more recent editorial piece, Chang reviewed the scope of the various automated delineation and segmentation methods. He reported that deep-learning-based approaches showed promise in areas of image recognition, object classification and disease detection, among its other advantages.4

Concerns and future outlook of AI-based tools & therapy 

Personalized medicine and patient care is today’s healthcare mantra, and possibly the way forward too. While artificial intelligence has definitely made a dramatic change to the way therapy can be approached, the role of radiation oncologists are still critical at this stage. The results from automated contouring data must be again verified by these specialists, section-by-section.8

Chang also reiterated this in his latest commentary: AI-based contouring could be a valuable tool in radiation therapy, but the algorithms need to be validated via more number of independent multicenter studies, larger patient cohorts, and they need to be proven clinically feasible before implementation.4,8 

A point, made by the investigators of the Lin L et al study, was concerning the development of better tools in the future and the inclusion of various other modalities within the same study. For example, their deep-learning algorithm was trained on an 3.0-T MRI data set, but whether the same would work with CT scan-data is unclear.7

However, in conclusion, there is no doubt that the application of AI in radiation oncology has been more than substantial—so far, process improvements and the reduction (and prevention) in error rates being the most significant. Looking ahead, the trend of machine learning and artificial intelligence, in general, could be the future of radiation oncology operations due to the precision and efficiency it demands. Though, it cannot be ignored that for AI to be embraced in health and medicine, governments need to act and put into place certain legislatures that legally protect institutional efforts.3

References 

  1. Thompson RF et al. Artificial intelligence in radiation oncology: A specialty-wide disruptive transformation? Radiotherapy and Oncology (2018), doi: https://doi.org/10.1016/j.radonc.2018.05.030
  2. World Economic Forum. The Fourth Industrial Revolution, by Klaus Schwab. https://www.weforum.org/about/the-fourth-industrial-revolution-by-klaus-schwab, Accessed May 22, 2019
  3. Weidlich V, Weidlich, GA. Artificial Intelligence in Medicine and Radiation Oncology. Cureus. 2018;10(4):e2475, doi: 10.7759/cureus.2475
  4. Chang Z. Will AI Improve Tumor Delineation Accuracy for Radiation Therapy? Radiology (2019), doi: https://doi.org/10.1148/radiol.2019190385
  5. Krizhevsky A, Sutskever I, Hinton, GE. ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems (2012), 1097-1105
  6. Tong N, Gou S, Yang S, Ruan D, Sheng K. Fully automatic multi‐organ segmentation for head and neck cancer radiotherapy using shape representation model constrained fully convolutional neural networks. Medical Physics (2018), doi: https://doi.org/10.1002/mp.13147
  7. Lin L et al. Deep Learning for Automated Contouring of Primary Tumor Volumes by MRI for Nasopharyngeal Carcinoma. Radiology (2019), doi: https://doi.org/10.1148/radiol.2019182012
  8. O’Connor M. Deep learning improves tumor contouring, may help patients with head and neck cancer, https://www.healthimaging.com/topics/artificial-intelligence/deep-learning-improves-tumor-contouring-help-cancer. Updated April 1, 2019. Accessed May 23, 2019
  9. Mark Klincewicz. Tumor Autocontouring — Efficiently Maximizing Dose and Minimizing Damage to Healthy Tissue. Radiology Today. 2009;10(11);22