Automatic Lung Segmentation and Quantification of Aeration in Computed Tomography of the Chest Using 3D Transfer Learning

Abstract

BackgroundIdentification of lung parenchyma on computer tomographic (CT) scans in the research setting is done semi-automatically and requires cumbersome manual correction. This is especially true in pathological conditions, hindering the clinical application of aeration compartment (AC) analysis. Deep learning based algorithms have lately been shown to be reliable and time-efficient in segmenting pathologic lungs. In this contribution, we thus propose a novel 3D transfer learning based approach to quantify lung volumes, aeration compartments and lung recruitability.MethodsTwo convolutional neural networks developed for biomedical image segmentation (uNet), with different resolutions and fields of view, were implemented using Matlab. Training and evaluation was done on 180 scans of 18 pigs in experimental ARDS (u2NetPig) and on a clinical data set of 150 scans from 58 ICU patients with lung conditions varying from healthy, to COPD, to ARDS and COVID-19 (u2NetHuman). One manual segmentations (MS) was available for each scan, being a consensus by two experts. Transfer learning was then applied to train u2NetPig on the clinical data set generating u2NetTransfer. General segmentation quality was quantified using the Jaccard index (JI) and the Boundary Function score (BF). The slope between JI or BF and relative volume of non-aerated compartment (SJI and SBF, respectively) was calculated over data sets to assess robustness toward non-aerated lung regions. Additionally, the relative volume of ACs and lung volumes (LV) were compared between automatic and MS.ResultsOn the experimental data set, u2NetPig resulted in JI = 0.892 [0.88 : 091] (median [inter-quartile range]), BF = 0.995 [0.98 : 1.0] and slopes SJI = −0.2 {95% conf. int. −0.23 : −0.16} and SBF = −0.1 {−0.5 : −0.06}. u2NetHuman showed similar performance compared to u2NetPig in JI, BF but with reduced robustness SJI = −0.29 {−0.36 : −0.22} and SBF = −0.43 {−0.54 : −0.31}. Transfer learning improved overall JI = 0.92 [0.88 : 0.94], P < 0.001, but reduced robustness SJI = −0.46 {−0.52 : −0.40}, and affected neither BF = 0.96 [0.91 : 0.98] nor SBF = −0.48 {−0.59 : −0.36}. u2NetTransfer improved JI compared to u2NetHuman in segmenting healthy (P = 0.008), ARDS (P < 0.001) and COPD (P = 0.004) patients but not in COVID-19 patients (P = 0.298). ACs and LV determined using u2NetTransfer segmentations exhibited < 5% volume difference compared to MS.ConclusionCompared to manual segmentations, automatic uNet based 3D lung segmentation provides acceptable quality for both clinical and scientific purposes in the quantification of lung volumes, aeration compartments, and recruitability.

Segmental Lung Recruitment in Patients with Bilateral COVID-19 Pneumonia Complicated by Acute Respiratory Distress Syndrome: A Case Report

“…Moreover, a 25–35% improvement in LUS score was observed in these patients within 48 h after the procedures, although it should be noted that the LUS score [ 18 ] is limited by the fact that the number of B-lines is influenced by many factors and should be used with caution [ 19 , 20 , 21 , 22 ]. Therefore, new tools, such as automatic deep learning-based algorithms, can be used to estimate whether a lung is recutable [ 23 ]. It is very important to emphasize that segmental lung recruitment had no effect on hemodynamic deterioration in patients, proving that segmental lung recruitment is a sparing and minimally invasive method [ 24 ].…”

Effects of Positive End-expiratory Pressure on Pulmonary Perfusion Distribution and Intrapulmonary Shunt during One-lung Ventilation in Pigs: A Randomized Crossover Study

“…Gas fractions were derived from CT values in Hounsfield units (HU) by Fgas=HU−1000. Lung segmentation was performed in a semiautomatic approach: (1) automatic segmentation of the end-inspiratory scan lung parenchyma by a deep convolutional neural network, as described in detail elsewhere 12 ; (2) manual refinement to include atelectatic lung regions (ITKSnap, Yushkevich et al . 13 ); (3) segmentation and exclusion of large airways up to the fifth generation by region-growing and local thresholding (3Dslicer, Fedorov et al .…”

Advanced Analysis of 3D Kinect Data: Supervised Classification of Facial Nerve Function via Parallel Convolutional Neural Networks

“…For example, a deep learning approach has been used in breast tumor histopathological images analysis [4,5], where the authors introduce a convolutional neural network to detect tumor targets from pathological images. Deep learning algorithms have recently been shown to be reliable and time-efficient in segmenting pathological lungs and quantification of aeration of the chest [6]. Deep neural networks also play an important role in data augmentation for brain tumor detection in magnetic resonance imaging [7].…”