Review ArticleContinuing Education

Intelligent Imaging: Anatomy of Machine Learning and Deep Learning

Geoff Currie
Journal of Nuclear Medicine Technology December 2019, 47 (4) 273-281; DOI: https://doi.org/10.2967/jnmt.119.232470

Article Figures & Data

Figures

  • FIGURE 1.
    FIGURE 1.

    The training phase of a neural network using extracted features as inputs. The grounded truth defines the ANN as a supervised ANN. This ANN structure is also what might be used as an analysis tool in parallel with traditional statistical analysis—importing data from a spreadsheet, for example. This example, which depicts all nodes as being connected to all others in adjacent layers, represents “fully connected layer,” which is more typical of CNNs. The validation phase evaluates trained ANN against a new database of known cases to determine accuracy. (Adapted from (1).)

  • FIGURE 2.
    FIGURE 2.

    Basic structure of a CNN, with the network extracting radiomic features, producing convolution function, pooling data through rectified linear unit (ReLU) kernel, and flattening pooled feature map for input into fully connected hidden layers of the neural network. (Reprinted from (1).)

  • FIGURE 3.
    FIGURE 3.

    Anatomy of an ANN. Single node (C) can have multiple inputs (X) with different weighting factors (W) and bias (B) but single output (Y) via activation function (A). Multiple lines exiting each node are same output being delivered to multiple next-layer nodes.

  • FIGURE 4.
    FIGURE 4.

    (A) Schematic representation of good fit vs. under- and overfitting associated with selection loss. (B, left) Optimization of selection loss to determine ANN complexity and node number (order) using decremental order algorithm, with arrow indicating reasonable cutoff for total node number. (B, right) Optimization of selection loss to determine inputs (features) to be included using growing-input algorithm, with arrow indicating reasonable cutoff for inputs.

  • FIGURE 5.
    FIGURE 5.

    Convolution uses 3 × 3 kernel to run sequential (in this case, successive, to provide stride of 1) 3 × 3 array of elements. Weighted sum of kernel for 3 × 3 input tensor creates single representative value in feature map. Multiple feature maps are produced by different kernels.

  • FIGURE 6.
    FIGURE 6.

    Pooling using max pooling method and 2 × 2 array produces pooling of maximum count among 4 connected elements (patch) to represent those data in pooled feature map. Use of consecutive blocks of 2 × 2 elements means stride of 2. Final pooled feature map immediately before input into neural network can then be flattened from 2-dimensional data into single dimension; this approach avoids need for global pooling.

  • FIGURE 7.
    FIGURE 7.

    CNN will have several convolution and pooling layers before flattening and input to neural network. Several kernels can be used on same input tensor to produce layers of feature maps via rectified linear units (ReLU) for pooling and eventually flattening.

Additional Files

Back to top

In this issue

Print
Download PDF
Article Alerts
Email Article
Citation Tools
Bookmark this article

Jump to section

Keywords