ShapeNet: Age-focused Landmark Shape Prediction with Regressive CNN
In this paper, we aim to predict the shape with inputting one single age.
We are the first to propose a regressive convolutional neural network named ShapeNet for landmark age-focused shape prediction. Unlike conventional CNNs, the input is only a single numeral (age), and the output is the corresponding shape instead of a label per pixel in a segmentation map. We demonstrate how to extend our ShapeNet into 3D space.
We validate our model using both synthetic data and human MRI and CT data. The results indicate that given a target age, ShapeNet is generally able to predict the shape with correct topology, and with much higher accuracy than a state-of-the-art geodesic approach.
The architecture of our proposed ShapeNet
An example of features from different layer using ShapeNet for predicting the shape of a pentagon at “age” 63.
Results
Shape prediction results for ages 1-100 of pentagon data, given data from ages 1-50 (GR: geodesic regression and LR: linear regression).
Shape prediction results of corpus callosum using ShapeNet and geodesic regression.
The predicted shapes of human mandible shapes with the increasing of age (as shown by the color).
The predicted shapes of human amygdala shapes with the increasing of age.