Lung Nodule Synthesis Using CNN-Based Latent Data Representation

Convolutional neural networks (CNNs) have been widely used to address various image analysis problems at the cost of intensive computational load and large amounts of annotated training data. When it comes to Medical Imaging, annotation is often complicated and/or expensive, and innovative methods for dealing with small or very imbalanced training sets are mostly welcome. In this context, this paper proposes a novel approach for efficiently synthesizing volumetric patch data from a small amount of samples using their latent data. Our method consists of two major steps. First, we train a 3D CNN auto-encoder for unsupervised learning of volumetric latent data by means of multivariate Gaussian mixture models (GMMs): while the encoder finds latent representations of volumes using GMMs, the decoder uses the estimated GMMs parameters to reconstruct the volume observed in the input. Then, we modify latent data of samples at training time to generate similar, but different, new samples: we run non-rigid registrations between patches decoded from real latent data and patches decoded from modified latent data, and warp the corresponding original image patches using the resulting displacement fields. We evaluated our method in the context of lung nodules synthesis using the publicly available LUNA challenge dataset, and generated new realistic samples out of real lung nodules, preserving their original texture and neighbouring anatomical structures. Our results demonstrate that 3D CNNs trained using our synthesis method were able to consistently deliver lower lung nodule false positive rates, which indicates an improvement in the networks discriminant power.