Fig 1.
Distributions of axial 18F-FDG PET intra slices extracted from the 3D tumor volume of non-responders and responders.
Fig 2.
Kaplan-Meier plot showing the survival rates of responders and non-responders.
Table 1.
Summary of second and high order texture features extracted from texture analysis.
Fig 3.
18F-FDG PET ROIs of a specific tumor i after segmentation embedded into larger square background of standard size of 100 × 100 pixels.
Each enlarged slice is denoted by xi,j and each set of three spatially adjacent enlarged slides is denoted by zi,k, where j and k represent the slices and triplets of the specific tumor i. In this example only 3 triplets, from the 5 available slices can be formed, so k = 1,2,3.
Fig 4.
CNN architecture for fusion of 3 adjacent 18F-FDG PET intra slices into a vector v.
The CNN architecture is composed from 4 convolutional and 4 max-pooling layers denoted by and
. In the first convolutional layer U(1), different coloured arrows represent the usage of different learnable weight matrices for convolving each PET slice in the triplet. Colored dotted rectangles in the feature maps represent elements of the feature maps that enclose local spatial information of the previous layer in the architecture. In the Max-pooling layers 2 × 2 element windows represent non-overlapped grids from which we choose the maximum element to downsample the feature maps.
Table 2.
Classification results: each figure is the average of three independent experiments using different training and test datasets.
Fig 5.
Ten most important texture features for prediction of the chemotherapy response using the GB algorithm.
Since these measures are relative, we assign the largest importance a value of 100% and then scale the others accordingly.
Fig 6.
Examples of feature maps in the first and last max-pooling layers V(1) , V(4) of the CNN architecture.
The feature maps illustrate how a specific triplet is represented in the first and last max-pooling layers.