Off-the-shelf 3D Lung Segmentation in CT using Generalized Histogram Thresholding

Abstract

Computerized Tomography is very important for lung disease diagnostics, including computer assisted methods. Lung segmentation is usually a first step in further sophisticated methods of diagnosis. If in one hand, deep learning methods have state-of-the-art performance, they aren't as simple to apply compared to classical methods, sometimes requiring extra data and training. We designed a method specific for lung segmentation based on histogram thresholding. We observed that, in our proposed method, by changing from Otsu to the more recently developed GHT we got a significant improvement in segmentation, jumping from 77% to 91% average dice (from 90% to 95% median dice, respectively), approaching deep learning methods (UNet) results (94% average and 97% median dice). Even though our proposed method runs on CPU, it's still 2.6 times faster than UNet on GPU. Moreover, our proposed method is off-the-shelf, requiring no training or parameter calibration, being suitable as pre-processing for more sophisticated methods that aim specific diagnoses.