Mathilde M. W. Wille

Pulmonary Nodule Detection in CT Images: False Positive Reduction Using Multi-View Convolutional Networks

We propose a novel Computer-Aided Detection (CAD) system for pulmonary nodules using multi-view convolutional networks (ConvNets), for which discriminative features are automatically learnt from the training data. The network is fed with nodule candidates obtained by combining three candidate detectors specifically designed for solid, subsolid, and large nodules. For each candidate, a set of 2-D patches from differently oriented planes is extracted. The proposed architecture comprises multiple streams of 2-D ConvNets, for which the outputs are combined using a dedicated fusion method to get the final classification. Data augmentation and dropout are applied to avoid overfitting. On 888 scans of the publicly available LIDC-IDRI dataset, our method reaches high detection sensitivities of 85.4% and 90.1% at 1 and 4 false positives per scan, respectively. An additional evaluation on independent datasets from the ANODE09 challenge and DLCST is performed. We showed that the proposed multi-view ConvNets is highly suited to be used for false positive reduction of a CAD system.

Results of the Randomized Danish Lung Cancer Screening Trial with Focus on High-Risk Profiling

RATIONALE As of April 2015, participants in the Danish Lung Cancer Screening Trial had been followed for at least 5 years since their last screening. OBJECTIVES Mortality, causes of death, and lung cancer findings are reported to explore the effect of computed tomography (CT) screening. METHODS A total of 4,104 participants aged 50-70 years at the time of inclusion and with a minimum 20 pack-years of smoking were randomized to have five annual low-dose CT scans (study group) or no screening (control group). MEASUREMENTS AND MAIN RESULTS Follow-up information regarding date and cause of death, lung cancer diagnosis, cancer stage, and histology was obtained from national registries. No differences between the two groups in lung cancer mortality (hazard ratio, 1.03; 95% confidence interval, 0.66-1.6; P = 0.888) or all-cause mortality (hazard ratio, 1.02; 95% confidence interval, 0.82-1.27; P = 0.867) were observed. More cancers were found in the screening group than in the no-screening group (100 vs. 53, respectively; P < 0.001), particularly adenocarcinomas (58 vs. 18, respectively; P < 0.001). More early-stage cancers (stages I and II, 54 vs. 10, respectively; P < 0.001) and stage IIIa cancers (15 vs. 3, respectively; P = 0.009) were found in the screening group than in the control group. Stage IV cancers were nonsignificantly more frequent in the control group than in the screening group (32 vs. 23, respectively; P = 0.278). For the highest-stage cancers (T4N3M1, 21 vs. 8, respectively; P = 0.025), this difference was statistically significant, indicating an absolute stage shift. Older participants, those with chronic obstructive pulmonary disease, and those with more than 35 pack-years of smoking had a significantly increased risk of death due to lung cancer, with nonsignificantly fewer deaths in the screening group. CONCLUSIONS No statistically significant effects of CT screening on lung cancer mortality were found, but the results of post hoc high-risk subgroup analyses showed nonsignificant trends that seem to be in good agreement with the results of the National Lung Screening Trial. Clinical trial registered with www.clinicaltrials.gov (NCT00496977).

Predictive Accuracy of the PanCan Lung Cancer Risk Prediction Model -External Validation based on CT from the Danish Lung Cancer Screening Trial

ObjectivesLung cancer risk models should be externally validated to test generalizability and clinical usefulness. The Danish Lung Cancer Screening Trial (DLCST) is a population-based prospective cohort study, used to assess the discriminative performances of the PanCan models.MethodsFrom the DLCST database, 1,152 nodules from 718 participants were included. Parsimonious and full PanCan risk prediction models were applied to DLCST data, and also coefficients of the model were recalculated using DLCST data. Receiver operating characteristics (ROC) curves and area under the curve (AUC) were used to evaluate risk discrimination.ResultsAUCs of 0.826–0.870 were found for DLCST data based on PanCan risk prediction models. In the DLCST, age and family history were significant predictors (p = 0.001 and p = 0.013). Female sex was not confirmed to be associated with higher risk of lung cancer; in fact opposing effects of sex were observed in the two cohorts. Thus, female sex appeared to lower the risk (p = 0.047 and p = 0.040) in the DLCST.ConclusionsHigh risk discrimination was validated in the DLCST cohort, mainly determined by nodule size. Age and family history of lung cancer were significant predictors and could be included in the parsimonious model. Sex appears to be a less useful predictor.Key points• High accuracy in logistic modelling for lung cancer risk stratification of nodules.• Lung cancer risk prediction is primarily based on size of pulmonary nodules.• Nodule spiculation, age and family history of lung cancer are significant predictors.• Sex does not appear to be a useful risk predictor.

Towards automatic pulmonary nodule management in lung cancer screening with deep learning

The introduction of lung cancer screening programs will produce an unprecedented amount of chest CT scans in the near future, which radiologists will have to read in order to decide on a patient follow-up strategy. According to the current guidelines, the workup of screen-detected nodules strongly relies on nodule size and nodule type. In this paper, we present a deep learning system based on multi-stream multi-scale convolutional networks, which automatically classifies all nodule types relevant for nodule workup. The system processes raw CT data containing a nodule without the need for any additional information such as nodule segmentation or nodule size and learns a representation of 3D data by analyzing an arbitrary number of 2D views of a given nodule. The deep learning system was trained with data from the Italian MILD screening trial and validated on an independent set of data from the Danish DLCST screening trial. We analyze the advantage of processing nodules at multiple scales with a multi-stream convolutional network architecture, and we show that the proposed deep learning system achieves performance at classifying nodule type that surpasses the one of classical machine learning approaches and is within the inter-observer variability among four experienced human observers.

Bag-of-Frequencies: A Descriptor of Pulmonary Nodules in Computed Tomography Images

We present a novel descriptor for the characterization of pulmonary nodules in computed tomography (CT) images. The descriptor encodes information on nodule morphology and has scale-invariant and rotation-invariant properties. Information on nodule morphology is captured by sampling intensity profiles along circular patterns on spherical surfaces centered on the nodule, in a multi-scale fashion. Each intensity profile is interpreted as a periodic signal, where the Fourier transform is applied, obtaining a spectrum. A library of spectra is created and labeled via unsupervised clustering, obtaining a Bag-of-Frequencies, which is used to assign each spectra a label. The descriptor is obtained as the histogram of labels along all the spheres. Additional contributions are a technique to estimate the nodule size, based on the sampling strategy, as well as a technique to choose the most informative plane to cut a 2-D view of the nodule in the 3-D image. We evaluate the descriptor on several nodule morphology classification problems, namely discrimination of nodules versus vascular structures and characterization of spiculation. We validate the descriptor on data from European screening trials NELSON and DLCST and we compare it with state-of-the-art approaches for 3-D shape description in medical imaging and computer vision, namely SPHARM and 3-D SIFT, outperforming them in all the considered experiments.

A Hierarchical Scheme for Geodesic Anatomical Labeling of Airway Trees

We present a fast and robust supervised algorithm for labeling anatomical airway trees, based on geodesic distances in a geometric tree-space. Possible branch label configurations for a given tree are evaluated based on distances to a training set of labeled trees. In tree-space, the tree topology and geometry change continuously, giving a natural way to automatically handle anatomical differences and noise. The algorithm is made efficient using a hierarchical approach, in which labels are assigned from the top down. We only use features of the airway centerline tree, which are relatively unaffected by pathology.

Visual assessment of early emphysema and interstitial abnormalities on CT is useful in lung cancer risk analysis

AbstractObjectivesScreening for lung cancer should be limited to a high-risk-population, and abnormalities in low-dose computed tomography (CT) screening images may be relevant for predicting the risk of lung cancer. Our aims were to compare the occurrence of visually detected emphysema and interstitial abnormalities in subjects with and without lung cancer in a screening population of smokers.MethodsLow-dose chest CT examinations (baseline and latest possible) of 1990 participants from The Danish Lung Cancer Screening Trial were independently evaluated by two observers who scored emphysema and interstitial abnormalities. Emphysema (lung density) was also measured quantitatively.ResultsEmphysema was seen more frequently and its extent was greater among participants with lung cancer on baseline (odds ratio (OR), 1.8, p = 0.017 and p = 0.002) and late examinations (OR 2.6, p < 0.001 and p < 0.001). No significant difference was found using quantitative measurements. Interstitial abnormalities were more common findings among participants with lung cancer (OR 5.1, p < 0.001 and OR 4.5, p < 0.001).There was no association between presence of emphysema and presence of interstitial abnormalities (OR 0.75, p = 0.499).ConclusionsEven early signs of emphysema and interstitial abnormalities are associated with lung cancer. Quantitative measurements of emphysema—regardless of type—do not show the same association.Key Points• Visually detected emphysema on CT is more frequent in individuals who develop lung cancer.• Emphysema grading is higher in those who develop lung cancer.• Interstitial abnormalities, including discrete changes, are associated with lung cancer. • Quantitative lung density measurements are not useful in lung cancer risk prediction. • Early CT signs of emphysema and interstitial abnormalities can predict future risk.

Optimal surface segmentation using flow lines to quantify airway abnormalities in chronic obstructive pulmonary disease

This paper introduces a graph construction method for multi-dimensional and multi-surface segmentation problems. Such problems can be solved by searching for the optimal separating surfaces given the space of graph columns defined by an initial coarse surface. Conventional straight graph columns are not well suited for surfaces with high curvature, we therefore propose to derive columns from properly generated, non-intersecting flow lines. This guarantees solutions that do not self-intersect. The method is applied to segment human airway walls in computed tomography images in three-dimensions. Phantom measurements show that the inner and outer radii are estimated with sub-voxel accuracy. Two-dimensional manually annotated cross-sectional images were used to compare the results with those of another recently published graph based method. The proposed approach had an average overlap of 89.3±5.8%, and was on average within 0.096±0.097mm of the manually annotated surfaces, which is significantly better than what the previously published approach achieved. A medical expert visually evaluated 499 randomly extracted cross-sectional images from 499 scans and preferred the proposed approach in 68.5%, the alternative approach in 11.2%, and in 20.3% no method was favoured. Airway abnormality measurements obtained with the method on 490 scan pairs from a lung cancer screening trial correlate significantly with lung function and are reproducible; repeat scan R(2) of measures of the airway lumen diameter and wall area percentage in the airways from generation 0 (trachea) to 5 range from 0.96 to 0.73.

Emphysema progression is visually detectable in low-dose CT in continuous but not in former smokers

ObjectivesTo evaluate interobserver agreement and time-trend in chest CT assessment of emphysema, airways, and interstitial abnormalities in a lung cancer screening cohort.MethodsVisual assessment of baseline and fifth-year examination of 1990 participants was performed independently by two observers. Results were standardised by means of an electronic score sheet; kappa and time-trend analyses were performed.ResultsInterobserver agreement was substantial in early emphysema diagnosis; highly significant (p < 0.001) time-trends in both emphysema presence and grading were found (higher prevalence and grade of emphysema in late CT examinations). Significant progression in emphysema was seen in continuous smokers, but not in former smokers. Agreement on centrilobular emphysema subtype was substantial; agreement on paraseptal subtype, moderate. Agreement on panlobular and mixed subtypes was only fair. Agreement was fair regarding airway analysis. Interstitial abnormalities were infrequent in the cohort, and agreement on these was fair to moderate. A highly significant time-trend was found regarding interstitial abnormalities, which were more frequent in late examinations.ConclusionsVisual scoring of chest CT is able to characterise the presence, pattern, and progression of early emphysema. Continuous smokers progress; former smokers do not.Key Points• Substantial interobserver consistency in determining early-stage emphysema in low-dose CT.• Longitudinal analyses show clear time-trends for emphysema presence and grading.• For continuous smokers, progression of emphysema was seen in all lung zones.• For former smokers, progression of emphysema was undetectable by visual assessment.• Onset and progression of interstitial abnormalities are visually detectable.

Malignancy risk estimation of screen-detected nodules at baseline CT: comparison of the PanCan model, Lung-RADS and NCCN guidelines

ObjectivesTo compare the PanCan model, Lung-RADS and the 1.2016 National Comprehensive Cancer Network (NCCN) guidelines for discriminating malignant from benign pulmonary nodules on baseline screening CT scans and the impact diameter measurement methods have on performances.MethodsFrom the Danish Lung Cancer Screening Trial database, 64 CTs with malignant nodules and 549 baseline CTs with benign nodules were included. Performance of the systems was evaluated applying the systems original diameter definitions: Dlongest-C (PanCan), DmeanAxial (NCCN), both obtained from axial sections, and Dmean3D (Lung-RADS). Subsequently all diameter definitions were applied uniformly to all systems. Areas under the ROC curves (AUC) were used to evaluate risk discrimination.ResultsPanCan performed superiorly to Lung-RADS and NCCN (AUC 0.874 vs. 0.813, p = 0.003; 0.874 vs. 0.836, p = 0.010), using the original diameter specifications. When uniformly applying Dlongest-C, Dmean3D and DmeanAxial, PanCan remained superior to Lung-RADS (p < 0.001 – p = 0.001) and NCCN (p < 0.001 – p = 0.016). Diameter definition significantly influenced NCCN’s performance with Dlongest-C being the worst (Dlongest-C vs. Dmean3D, p = 0.005; Dlongest-C vs. DmeanAxial, p = 0.016).ConclusionsWithout follow-up information, the PanCan model performs significantly superiorly to Lung-RADS and the 1.2016 NCCN guidelines for discriminating benign from malignant nodules. The NCCN guidelines are most sensitive to nodule size definition.Key Points• PanCan model outperforms Lung-RADS and 1.2016 NCCN guidelines in identifying malignant pulmonary nodules.• Nodule size definition had no significant impact on Lung-RADS and PanCan model.• 1.2016 NCCN guidelines were significantly superior when using mean diameter to longest diameter.• Longest diameter achieved lowest performance for all models.• Mean diameter performed equivalently when derived from axial sections and from volumetry.