Abstracts

RETURN TO ABSTRACT LISTING


E1609. Convolutional Neural Networks Made Simple: A Concise Tutorial For Radiologists
Authors
  1. Jeremy Nguyen; Tulane University School of Medicine
  2. Quan Nguyen; West Chester Medical Center
  3. Thao-Quyen Ho; University Medical Center - Ho Chi Minh City
  4. Mandy Weidenhaft; Tulane University School of Medicine
Background
A convolutional neural network (CNN) is a class of deep learning neural networks, which have extensive applications in radiologic imaging. CNN is constructed with artificial neurons connected together to form building blocks which include an input layer, hidden layers and an output layer. CNNs employ a mathematical operation called convolution for extracting features from the input image. CNNs can take in an input image, assign importance to various aspects/features in the image and be able to differentiate one from the other. This exhibit will provide a concise tutorial of CNNs without complex mathematics. The radiologist will learn the fundamental architecture of a Convolutional Neural Network (CNN). The concept of mathematical convolution for image feature extraction will be intuitively explained. The radiologist will learn the conceptual design of CNN with the integration of the feature extraction and image classification components. A self-assessment quiz is available at the end for assessment of material comprehension.

Educational Goals / Teaching Points
•To describe the structure and operation of an “artificial neuron” in a neural network. •To discuss the basic architecture of a deep learning neural network including the input, hidden and output layers. •To describe the architecture of a Convolutional Neural Network (CNN). •To explain the intuitive meaning of mathematical convolution. •To discuss how a convolutional neural network can extract features of an image though the Convolution Layer, Rectified Linear Unit Layer (ReLU), and Pooling. •To explain how a convolutional neural network can classify an image through the Fully Connected Layer and Softmax function. •To describe an intuitive design of a Convolutional Neural Network, with the integration of the feature extraction and image classification components.

Key Anatomic/Physiologic Issues and Imaging Findings/Techniques
• Deep learning • Convolutional neural network • Convolution Layer, Rectified Linear Unit Layer (ReLU), and Pooling for image feature extraction. • Fully Connected Layer and Softmax function for image classification

Conclusion
Convolutional neural network (CNN) has become a dominant artificial intelligence technology in radiology. CNN is constructed to automatically and adaptively learn the spatial hierarchies of features by using image feature extraction of classification building blocks. CNN can perform radiologic task such as classification, segmentation, and detection. After the completion of this tutorial, the radiologist will have a firm conceptual knowledge of CNN; giving the radiologist a firm foundation to further explore advanced applications of CNN in radiologic imaging