Chest X-Ray Pneumothorax Segmentation with the Multistep Post-Processing

Vladimir GROZA, Artur KUZIN

Keywords: Neural Network, Deep Learning, Segmentation, Medical Imaging

TL;DR: TLDR

Abstract:

In this work we present recent results of the pneumothorax segmentation from the chest X-ray images. Pneumothorax may appear in case of dull chest injury, as a continuation of hidden problems with the lungs, or even more there could be no reason at all for finding. In several situations, lung collapse can turn out as serious threat to life.
We propose new method which includes the chest X-ray image segmentation pipeline with the multistep conditioned post-processing. As the result, we demonstrate significant improvement compare to any strong ”baseline” by reduction of the pneumothorax collapse regions which are missed out and of false positive detections. Our results indicate very high accuracy and strong robustness of the algorithm confirmed by corresponding efficiency on the two stage test dataset with a priori unknown and absolutely different distribution. Final Dice scores 0.8821 and 0.8614 for ”stage 1” and ”stage 2” test sets respectively were resulted in top 0.01% standing of the private leaderboard on Kaggle competition platform.

Track: short paper

Paper Type: both