A radiologist’s guide to deep learning
Deep Learning (DL) has emerged as a powerful tool that has the potential to become an ally for radiologists in executing decision-making tasks.
Eight members of the Association of University Radiologists Radiology Research Alliance Task Force on Deep Learning published an article online March 30 in Academic Radiology. In it, they describe a radiologist-friendly overview examining past, present and future applications, while investigating how the field may benefit from embracing DL.
Below are takeaways from their research.
Neural Networks and Radiology
Deep learning is not a single algorithm—rather it is a method involving layered information inputs that make up a complex algorithm of networks called a convolutional neural networks (CNNs). Authors compared its structure to the biological nervous system.
CNNs, which specialize in image segmentation and classification, are the most applicable form of DL to radiology. However, the challenge has been the need for mass amounts of computing power, which forces researchers to limit the learning of an algorithm in specific, predetermined ways.
“The goal of DL is an intelligent computer system that can disentangle massive amounts of unlabeled and unstructured data to produce complex and meaningful insights,” wrote corresponding author William F. Auffermann, MD, PhD, with the department of radiology and imaging sciences at Emory University School of Medicine in Atlanta, and colleagues. “In a sense, this is the workflow of the modern radiologist—translating a large, digital dataset containing pixel intensity values into an accurate diagnosis.”
Current Applications in Radiology
“Radiology differs from other image recognition applications of DL algorithms in that a computed tomography or magnetic resonance imaging examination can consist of thousands of images as opposed to a single image,” Aufferman et al. wrote. “This greatly increases the complexity of required computational algorithms.”
Many uses of DL focus on a specific task such as lesion or disease detection, classification, diagnosis, segmentation and quantification.
DL systems utilizing nonmedical data have demonstrated effectiveness in categorizing findings on chest radiographs, such as pleural effusion, cardiomegaly and mediastinal enlargement.
A separate recently published study cited by Aufferman et al. used a pre-trained DL system to classify tuberculosis on chest radiographs with an area under the curve of 0.99. A subsequent radiologist-augmented approach achieved a sensitivity of 97.3 percent and specificity of 100 percent.
Similarly, deep-learning systems have shown promise in identifying pulmonary nodules and characterizing imaging abnormalities, potentially freeing up time spent by radiologists meticulously searching for small lesions.
While traditional machine-learning classification requires predefined features, DL algorithms can create their own, and several studies have accurately demonstrated the ability of CNNs to classify lung nodules as benign or malignant.
Further studies have shown success in CNNs’ ability to classify breast density and tumors in mammography and determine bone age in musculoskeletal imaging.
Radiomics has undergone a “rapid revolution” since the introduction of machine-learning techniques, Aufferman et al. wrote. These new approaches can extract large sets of medical imaging features, unseen by human eyes to correlate imaging features with clinical factors, diagnosis and outcomes.
Whereas brain MRI segmentation was once a tedious task, DL algorithms have proven effective at manually achieving the job.
Aufferman et al. point to a recent study which achieved accurate automatic organ segmentation with CNNs trained only on a single manually segmented image.
While limitations remain, the future is bright
For all the remarkable advances in deep learning and its potential applications, these neural networks are far from perfect and require further research.
Reverse engineering systems can alter input data invisible to the naked eye, and ethical and legal challenges linger around who is liable when an algorithm commits an error or misinterprets an image.
“Although great promise has been shown with DL algorithms in a variety of tasks across radiology and medicine as a whole, these systems are far from perfect,” according to Aufferman et al. “Neural networks can be ‘statistically impressive, but individually unreliable’ and can make mistakes that humans would not.”
However, DL has many potential future applications in radiology—including worklist optimization, diagnostic applications, prognostication, tracking automated image findings and automated preliminary report generation.
“DL offers exciting opportunities for radiologists to improve safety by providing more accurate diagnoses, increasing efficiency by automating tasks, and helping to generate data on imaging features that were not previously used as diagnostic criteria,” Aufferman et al. wrote.
Matt joined Chicago’s TriMed team in 2018 covering all areas of health imaging after two years reporting on the hospital field. He holds a bachelor’s in English from UIC, and enjoys a good cup of coffee and an interesting documentary.