False-Positive Reduction on Lung Nodules Detection in Chest Radiographs by Ensemble of Convolutional Neural Networks

Aiming at the problem that traditional lung nodules detection method can only get low sensitivities with a lot of false positives, we propose a new framework of ensemble of convolutional neural networks (E-CNNs) and use it to significantly reduce the number of false positive on lung nodules detection in chest radiographs (CXRs). First, unsharp mask technique is used to enhance the nodules in the CXRs. Then, we cut patches in the <tex-math notation="LaTeX">229×229 </tex-math> image containing or not containing nodule from the enhanced CXRs, which correspond to the positive and negative samples. Third, three optimized CNNs of different input sizes and different depths, namely, CNN1, CNN2 and CNN3, are constructed to detect lung nodule separately, and their input sizes are <tex-math notation="LaTeX">12×12 </tex-math> , <tex-math notation="LaTeX">32×32 </tex-math> , and <tex-math notation="LaTeX">60×60 </tex-math> , and the number of layers are 5, 7, and 9, separately. Finally, a logical AND operator is used to fuse the results of CNN1, CNN2, and CNN3, and E-CNNs are constructed for detecting lung nodules. Our experimental results on the Japanese Society of Radiological Technology database show our proposed E-CNNs attain a sensitivity of 94% and 84% with an average of 5.0 false-positives (FPs) per image and 2.0 FPs per image, respectively, in a five-fold cross-validation test, which far outperforms the state of the art.