The use of computer-aided secondary tools that assist physicians for automatic segmentation of the lung is very important for the diagnosis of many lung diseases. Performing the segmentation of the lungs with manual selection by experts can cause human error, subjective and unnecessary waste of time. In this study, an approach based on Mask R-CNN is proposed for automatic segmentation of the lung from CT scans. The CT image series used in the study are obtained from publicly available datasets such as HUG-ILD and VESSEL12 databases. In experimental studies for lung segmentation, for the HUG-ILD dataset, 95.95% Dice similarity coefficient and 7.65% volumetric overlap error are obtained, respectively, whereas for the VESSEL12 dataset, these metrics are measured as 96.80% and 6.12%, respectively. As a result, the Mask R-CNN-based approach proposed in this study for lung segmentation is confirmed to be successful.
Paper
Full text
A Mask R-CNN based Approach for Automatic Lung Segmentation in Computed Tomography Scans
Semantic Scholar · Medicine · 2021
Abstract
The use of computer-aided secondary tools that assist physicians for automatic segmentation of the lung is very important for the diagnosis of many lung diseases. Performing the segmentation of the lungs with manual selection by experts can cause human error, subjective and unnecessary waste of time. In this study, an approach based on Mask R-CNN is proposed for automatic segmentation of the lung from CT scans. The CT image series used in the study are obtained from publicly available datasets such as HUG-ILD and VESSEL12 databases. In experimental studies for lung segmentation, for the HUG-ILD dataset, 95.95% Dice similarity coefficient and 7.65% volumetric overlap error are obtained, respectively, whereas for the VESSEL12 dataset, these metrics are measured as 96.80% and 6.12%, respectively. As a result, the Mask R-CNN-based approach proposed in this study for lung segmentation is confirmed to be successful.