An efficient multi-scale data representation method for lung nodule false positive reduction using convolutional neural networks

In the last few years, convolutional neural networks (CNN) have been widely used to address very different image analysis problems. In medical imaging, they are becoming standard for aid in diagnosis in many imaging modalities at the cost of great computational load and huge amounts of annotated training data. When it comes to 3D image exams, the computational cost can easily became prohibitive demanding methods that promote efficiency in data representation. In this context, this paper presents a novel sampling method based on trigonometric functions that delivers comprehensive bidimensional data representation out of 3D volumetric patches for an efficient CNN-based classification. We used the publicly available LUNA challenge dataset and demonstrate that our representation allows 2D CNNs to achieve very good results, superior to the ones delivered using regular 2D cross-sections and similar to the ones delivered by 3D CNNs using 16 times less data and running 4 times faster.