Room: AAPM ePoster Library
Purpose: to test the ability of conventional convolutional neural net (CNN) architectures to effectively capture tumor shape properties, which are important predictors of cancer outcome.
Methods: A set of 5000 synthetic PET images was generated (voxel size 1.5 mm) containing tumors 20-60 mm in size. Image noise was added and a 3.0 mm FWHM Gaussian filter was applied to simulate resolution blurring. Shape features were computed for all images, including: volume; surface area; convex volume; solidity as the ratio of volume and convex volume; extent as the ratio of lesion volume to that of the lesion bounding box; and compactness. A conventional CNN with 7 convolutional layers, followed by 2 fully-connected layers was trained to predict the values of shape features from the images. The network was tested on a distinct set of 100 images. On the test set, we measured the percent mean absolute error (MAE) between the predicted and true feature values.
Results: The test MAEs varied greatly between the features. Size-related features were predicted with high accuracy: the MAE was 2.5% for volume, and 4.3% for surface area, and 5.7% for convex volume. On the other hand, features related to shape irregularity could not be effectively predicted by the CNN: the MAE was 17.7% for extent, 15.5% for solidity, and 16.4% for compactness. The high prediction errors were verified by inspecting the predicated-vs-actual scatter plots. The poor prediction accuracy is likely to be related to the ratio-type functional form of the shape irregularity features.
Conclusion: Radiomic features that describe shape irregularity cannot be effectively captured by conventional CNN architectures. In clinical prediction tasks, such features should be added to CNNs as auxiliary variables. Alternatively, standard CNN architectures can be modified to account for ratio-type functions. This finding bears significant implications for CNN-based clinical outcome prediction in radiology.