3.1.

Lobe Segmentation

Characterizing emphysema across a large lung screening cohort requires automatic lobe segmentation that is robust to smoking-related pathology. The leading algorithm from Hofmanninger et al.³³ has been validated on diverse lung pathology but exhibits substantial artifact at the pulmonary fissures (Fig. 2) that are not consistent with realistic 3D anatomy. To overcome this challenge, we employed a volumetric level set method³⁴ (LSM) to infer smoothly contoured and artifact-free borders [Fig. 3(b)] from the leading algorithm. LSM-evolved segmentations were acquired for each scan in VLSP and used as pseudo-labels to train a 3D U-Net from random weights [Fig. 3(c)]. The resulting model was then finetuned with near ground-truth annotations from TotalSeg [Fig. 3(d)]. The final model was quantitatively validated on an external screening CT dataset and qualitatively reviewed for a sample hold out set of VLSP [Fig. 3(e)].

Fig. 2

State-of-the-art lobe segmentation algorithms fail to capture realistic lobe fissures in 3D in lung screening CT. The leading lobe segmentation model from Hofmanninger et al.³³ applied to two examples. An irregular artifact is noticeable along the borders of the segmentation. In addition, the right middle lobe (orange) and right lower lobe (red) are inaccurately segmented.

Fig. 3

Overview of methods. (a) The 2D U-Net based segmentation algorithm from Hofmanninger et al.³³ is applied to infer initial segmentations for the VLSP cohort. Pulmonary lobe annotations are absent for this cohort. (b) Initial lobe segmentations are iteratively evolved using a volumetric level set method (LSM). (c) The evolved segmentations are used as pseudo-labels for VLSP in the self-supervised training of our 3D U-Net. (d) The model is further finetuned with 660 annotated chest CTs from the TotalSegmentator dataset. (e) An external validation set from LUNA16 is used to test the model’s performance on lung screening CTs. (f) The final model is leveraged to anatomically quantify emphysema severity for the NLST cohort. (g) Multivariate logistic regression is applied to study the association between emphysema measures and lung cancer incidence.

3.2.

Finetune Training and Validation

The model was finetuned with 660 CT scans and corresponding labels from TotalSegmentator using the same settings for data augmentation and loss criterion. Finetune training continued until the running average of the validation Dice decreased 10% from its peak. The finetuned model was tested using an external validation set consisting of 47 lung screening chest CTs from LUNA16 with corresponding radiologist-generated labels. During inference, our algorithm found connected components and set missing labels to their nearest neighbors as post-processing of U-Net outputs. For comparison, we computed the Dice score of original and LSM-corrected segmentations from Hofmanninger et al.³³ We report Dice scores for each pulmonary lobe as well as the average across the validation set (Fig. 4).

Fig. 4

LAV fraction across pulmonary lobes for lung cancer cases and controls without lung cancer. Greater variance in LAV fraction was observed in the RUL compared with other lobes.

3.3.

Emphysema Characterization

We investigated the association between lobar emphysema and lung cancer in a lung screening cohort sampled from NLST. Emphysema was approximated as the volume below the standard low attenuation threshold of −950 HU. The LAV was intersected with labels inferred from our lobe segmentation model to quantify the LAV percentage in each lobe. Whole lung volume was computed by multiplying the voxel count with the voxel dimensions.

3.4.

Statistical Analysis

We adopted the practice of Labaki et al.¹⁶ in adjusting for covariates from the ${\mathrm{PLCO}}_{\mathrm{m}2012}$⁶ lung cancer risk model, which were found to be significant predictors of developing lung cancer. These include age, sex, body mass index (BMI), COPD status, past cancer history, past lung cancer family history, smoking status, smoking quit time, and smoking pack years. We inferred missing values of BMI and/or COPD status for eight subjects using multivariate imputation. Continuous covariates were min-max normalized with the following equation:

Table 4

Adjusted odds ratios for lobar LAV with respect to lung cancer risk.

Table 5

Added value of lobar LAV to goodness-of-fit.

Dice performance of lobe segmentation is reported as the mean from the LUNA16 test set. 95% confidence intervals for the mean Dice were computed from 100 bootstrap samples by sampling with replacement from the predictions of LUNA16. We used a two-tailed Wilcoxon signed rank test to determine if there was a significant difference in Dice performance between our approach and baselines with a significance level of $\mathrm{p}<0.05$.

4.1.

Emphysema Characterization

4.2.

Lobe Segmentation

4.2.1.

Quantitative validation

There was no significant difference in Dice performance between the baseline and LSM-processed results. In contrast, our proposed method (0.9278 Dice) significantly outperformed the baseline (0.8996 Dice) and LSM (0.8982 Dice). Across each pulmonary lobe, the baseline and LSM achieved a similar Dice performance, whereas our proposed method achieved a significantly higher average Dice. A two-tailed Wilcoxon signed rank test found these differences to be significant for $\mathrm{p}<0.05$. All three methods performed worse on the RML compared with the other lobes, but here the improvement from our proposed method was pronounced (Fig. 5).

Fig. 5

Mean Dice similarity coefficient across all cross validation folds on LUNA16. “Hof. et al.” is the algorithm from Hofmanninger et al.³³ The LSM is evolved from Hof. et al using level set regularization and connected components. Our proposed model is trained using self-supervision and further fine-tuned on five cross validation folds with the combined LUNA16 and active learning dataset. † $\mathrm{p}<0.05$.

4.2.2.

Qualitative sensitivity analysis

No significant difference in the mean score between cancer status, COPD status, and sex were observed (Table 6). The mean scores for all clinical categories were above 4. Across all three categories, the score of 5 was the most frequent, reflecting the observation that these segmentations appeared to be accurate without any noticeable artifact. Qualitative performance was limited by a fewer number of 4s and 3s. These segmentations were ones that appeared less accurate and/or had some noticeable artifact, most commonly in subjects with challenging RML anatomy and obscure lobe fissures. A small number of scans received a score of 2, and no scans received a score of 1 (Fig. 6).

Table 6

Qualitative scores by cancer, COPD, and sex.

	Mean score	Std	P
Cancer present	4.45	0.72	<0.05
Cancer absent	4.25	0.85
COPD present	4.16	0.87	<0.05
COPD absent	4.34	0.80
Male	4.03	0.86	<0.05
Female	4.51	0.73

Fig. 6

Distributions of qualitative scores for a random sample of VLSP. The distributions of scores are heavily skewed toward a score of 4 and 5 regardless of lung cancer status, COPD status, and sex.

5.1.

Emphysema Characterization

We investigated lobar emphysema as an independent risk factor for lung cancer by applying an automated pipeline to quantify emphysema markers for a large lung screening cohort. Based on the quantitative percentage of LAV, we found higher emphysema involvement in the upper lobes, especially the RUL, compared with the lower lobes. These findings agree with previous studies of upper lobe-predominant emphysema being highly prevalent in lung screening cohorts.^9,16,40 The finding that LAV in the RUL independently contributes to lung cancer risk is novel. Our findings remain significant even when controlling for whole lung emphysema diagnosis, reinforcing the independent contribution of regional emphysema on lung cancer risk. Our results align with a previous study²² that found upper lobe emphysema to be a significant predictor of lung cancer location, although the distinction between left and right upper lobes was not made in that study. Our results also agree with a pooled meta-analysis of three studies^9,17,21 showing that centrilobular emphysema, which presents with upper lobe predominance, is independently associated with increased lung cancer risk. We did not confirm whether centrilobular emphysema was driving the observed association in our cohort because diagnosing emphysema subtype requires a visual assessment.

The lack of significant association between the LUL LAV percentage and lung cancer in our cohort is surprising given that the upper lobes are thought to be similarly affected in emphysema. The large odds ratio of the RLL LAV percentage with a 0.395 p-value is also worth noting and suggests a distinction between left and right lungs that has not been studied in the context of emphysema and lung cancer. The right main bronchus is wider, shorter, and more vertical than the left. The emphysema quantification of the LUL combines the LUL and the lingula, which is anatomically like the RML. However, the left and right upper lungs are thought to be physiologically similar in ventilation and perfusion despite these anatomical differences. Evidence of distinction between left and right lungs is more likely specific to this study’s cohort, and validation in other cohorts is needed for further investigation.

This section of work is limited in several aspects. First, Caucasians are overrepresented in both cases and controls of the NLST cohort. This may bias the results and limit their applicability to populations of greater racial diversity. Second, inaccurate segmentations may introduce noise into the quantitative emphysema measures, especially when quantifying LAV in the RML. We strived to minimize this noise by optimizing our model for lung screening CT and standardizing the analysis to soft kernel CTs. Third, large emphysematous bullae may confound emphysema quantification as these structures are known to compress surrounding lung tissue. This would nonlinearly influence LAV-based quantification.

5.2.

Lobe Segmentation

The present work seeks to develop a lobe segmentation algorithm that appears realistic in 3D and is robust to smoking-related changes in the lung. To this end, we proposed a two-stage training strategy that combines self-supervised training on a lung screening dataset with level set regularization and finetuning on a clinical routine dataset with near ground truth labels. In general, the LSM was successful in smoothing local artifacts, such as those in subjects A and B in Fig. 7, but it preserved large scale defects such as those in subjects C. Although such defects introduced label noise into the first stage of training, the final model resolved both small- and large-scale artifacts. Testing on an external validation set confirmed that our algorithm outperformed a leading baseline method on lung screening subjects across all lung lobes. As the most variable in anatomical shape and position, the RML was unsurprisingly the most challenging to segment. Although superior to the leading method, our algorithm was only moderately accurate in segmenting the RML.

Fig. 7

Lobe segmentations and ground truth annotations for three subjects using three methods: a pretrained baseline model, the baseline model post-processed with the LSM, and our proposed model. Segmentations of subject A, B, and C come from the 75’th, 50’th, and 25’th percentiles of our proposed model’s Dice performance. The LSM resolves small-scale border artifacts and achieves smoothly contoured fissures, but large-scale defects such as those seen in subject C’s RML remain unresolved. Our proposed method effectively resolves large-scale defects and segments the RML more accurately than the baseline method.

Our algorithm was not free of errors, as suggested by a mix of scores in our sensitivity analysis. Common failure modes included inaccurate segmentation of challenging RML anatomy and presence of visually apparent artifacts (Fig. 7). The scores were calibrated such that a 3 or above would be acceptable for use in most visualization applications. Because the overwhelming majority of samples scored in the acceptable range, this analysis supported the clinical utility of our algorithm. Because our algorithm was primarily developed and validated on lung screening subjects who all have substantial smoking history, its performance on pulmonary pathologies seen in the clinical routine is unknown.