Artificial IntelligenceUnderstanding artificial intelligence based radiology studies: CNN architecture
Introduction
Artificial intelligence (AI) is a broad umbrella term used to encompass a wide variety of subfields dedicated to creating algorithms to perform tasks that mimic human intelligence. Deep learning is a type of machine learning inspired by the biological nervous system that uses multiple layers of simple processing units called nodes that are intricately interconnected. Similar to the function of nodes in the nervous system, deep learning algorithms automatically extract and combine progressively higher-level representations of an input in order to perform the task required. There is excitement around deep learning's ability to discover previously unknown relationships in data and perform complex tasks. While deep learning is conceptually not a new idea,1 the application of this technology became a reality in the late 2000's with the advancements in graphics processing units (GPUs), which enabled complex computational requirements of multiple matrix multiplication with parallel processing.
In 2012, another breakthrough came in the form of a new convolutional neural network (CNN) introduced at the annual ImageNet challenge competition. This CNN, named the AlexNet,3 used an innovative architecture to achieve a performance significantly better than previous years. AlexNet provided a new insight into the framework of how neural networks can be developed and became a foundation for many of the currently used networks. Later in 2017, Google released tensor processing units (TPUs) that offered even more computational power.2 This resulted in lower cost and greater accessibility to these hardware developments, enabling further progress in the field of deep learning.
This technology has gained interest in the field of medical imaging and has been applied to tumor detection, pathologic correlation, and treatment response monitoring.[4], [5], [6],7., [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18] Preliminary studies have yielded promising results with performance in some detection tasks comparable to a trained radiologists.19 In this article, we will review general principles of neural network architecture and its relevance towards radiologic image analysis.
Section snippets
Convolutional neural network architecture basics and terms
Neural networks were inspired by the biological nervous system composed of neuronal connections. These networks were further developed mimicking the functionality of the human visual system20,21 which operates by progressively combining the input of neurons that recognize simple features such as lines or colors in the visual field into more complex representations. Fig. 1 illustrates the general foundation of a CNN architecture.
Review of architectures
Many types of CNN architectures have been developed that are tailored for different tasks including classification, detection, and segmentation.
Conclusion
Deep learning technology has made remarkable progress in the last decade and many medical studies have found success using neural networks to perform classification, detection, and segmentation. Understanding basics concepts of neural network architecture and its application in imaging analysis will be essential for radiologists as AI based software programs become more integrated into the clinical workflow.
Declaration of competing interest
No conflict of interest.
References (63)
- et al.
Artificial intelligence and machine learning in radiology: opportunities, challenges, pitfalls, and criteria for success
J Am Coll Radiol
(2018 Mar) - et al.
Large scale deep learning for computer aided detection of mammographic lesions
Med Image Anal
(2017) - et al.
Understanding artificial intelligence based radiology studies: what is overfitting?
Clin Imaging
(2020) - et al.
3D fully convolutional networks for subcortical segmentation in MRI: a large-scale study
Neuroimage
(2018) - et al.
Cybernetics and forecasting techniques
(1967) - et al.
In-datacenter performance analysis of a tensor processing unit
- et al.
ImageNet classification with deep convolutional neural networks
- et al.
A roadmap for foundational research on artificial intelligence in medical imaging: from the 2018 NIH/RSNA/ACR/The Academy Workshop
Radiology
(2019) - et al.
Artificial intelligence in radiology
Nat Rev Cancer
(2018) - Huang X, Shan J, Vaidya V. Lung nodule detection in CT using 3D convolutional neural networks. 2017 IEEE 14th...