integrityNet
Zachary W. Althof, Joseph M. Reinhardt, Sajan Goud Lingala, Osama I. Saba, Gary E. Christensen, Eric A. Hoffman · 2021
Lungs are regularly imaged through Computed Tomography (CT). The submillimeter resolution allows for intricate structures and features within the lungs to be visualized and analyzed. Quantitative data and region segmentation from CT images is essential to efficiently process data in clinical and research settings. Deep neural networks have been shown to be able to effectively automate difficult and complex diagnoses and segmentations from medical images. Within the lungs there are separate regions, called lobes, separated by surfaces, called fissures. Fissures can completely or partially separate the lobes. Regions of incompleteness are known to be a biomarker of collateral ventilation between lobes which is clinically relevant for endobronchial valve treatment patient screening. Beyond this, the extent and location of fissure completeness may impact lung biomechanics and have implications in disease susceptibility and progression. Automated quantification and localization of fissure completeness, or integrity, is needed to study these areas of research in large datasets. In this work, a deep learning technique is utilized to automate fissure integrity segmentation. Using a state-of-the-art fissure segmention technique to provide input localization information, the network was able to leverage the predictive power of the fissure segmentation method to improve the model’s performance. Furthermore, the U-Net structure of the proposed network allows for better spatial prediction through the use of high-level features during image reconstruction.The model was trained and evaluated using a radiologist trained hand segmented imaging dataset. Two tasks were assessed: 1) fissure integrity classification into different categories defined by ranges of fissure completeness percentages and 2) fissure integrity localization to produce images labeling regions of incompleteness along the fissure surface. The results demonstrate that the network was able to effectively quantify fissure integrity percentage and produce fissure integrity images similar to the ground truth images in the dataset. The model achieved high accuracy in both tasks and has potential to be used in widescale investigation of fissure integrity trends and relations to biomechanics and disease.