Artificially-generated consolidations and balanced augmentation increase performance of U-net for lung parenchyma segmentation on MR images

PURPOSE: To improve automated lung segmentation on 2D lung MR images using balanced augmentation and artificially-generated consolidations for training of a convolutional neural network (CNN). MATERIALS AND METHODS: From 233 healthy volunteers and 100 patients, 1891 coronal MR images were acquired. Of these, 1666 images without consolidations were used to build a binary semantic CNN for lung segmentation and 225 images (187 without consolidations, 38 with consolidations) were used for testing. To increase CNN performance of segmenting lung parenchyma with consolidations, balanced augmentation was performed and artificially-generated consolidations were added to all training images. The proposed CNN (CNN(Bal/Cons)) was compared to two other CNNs: CNN(Unbal/NoCons)—without balanced augmentation and artificially-generated consolidations and CNN(Bal/NoCons)—with balanced augmentation but without artificially-generated consolidations. Segmentation results were assessed using Sørensen-Dice coefficient (SDC) and Hausdorff distance coefficient. RESULTS: Regarding the 187 MR test images without consolidations, the mean SDC of CNN(Unbal/NoCons) (92.1 ± 6% (mean ± standard deviation)) was significantly lower compared to CNN(Bal/NoCons) (94.0 ± 5.3%, P = 0.0013) and CNN(Bal/Cons) (94.3 ± 4.1%, P = 0.0001). No significant difference was found between SDC of CNN(Bal/Cons) and CNN(Bal/NoCons) (P = 0.54). For the 38 MR test images with consolidations, SDC of CNN(Unbal/NoCons) (89.0 ± 7.1%) was not significantly different compared to CNN(Bal/NoCons) (90.2 ± 9.4%, P = 0.53). SDC of CNN(Bal/Cons) (94.3 ± 3.7%) was significantly higher compared to CNN(Bal/NoCons) (P = 0.0146) and CNN(Unbal/NoCons) (P = 0.001). CONCLUSIONS: Expanding training datasets via balanced augmentation and artificially-generated consolidations improved the accuracy of CNN(Bal/Cons), especially in datasets with parenchymal consolidations. This is an important step towards a robust automated postprocessing of lung MRI datasets in clinical routine.