Patient Characteristics
Among the lung cancer patients in our previous study (n = 137, ClinicalTrials.gov, NCT03648151) [11], 117 patients had primary tumor and more than one lymph node, simultaneously. Furthermore, among the analyzed lymph nodes of the previous cohort (n = 1239), SUVmax of 413 and 322 ones were higher than 2.0 and 2.5, respectively. The differential diagnosis of lymph nodes is illustrated on Fig. 1. Finally, for the cohort 1 and 2, treatment modalities of 94 and 88 patients could be followed, and median number of malignant foci was 4 (ranged from 2 to 16) and 3.5 (ranged from 2 to 15), respectively. Additionally, median age of cohort 1 was 65y (ranged from 38y to 88y), and was used to group the patients. In general, patient characteristics were similar between the two cohorts (Table 1), and indicated that the patient stratifications could not be obviously influenced by the thresholds of SUVmax.
Table 1
Patient characteristics of the cohorts. SUVmax of lymph nodes in cohort 1 (n = 94) and 2 (n = 88) is higher than 2.0 and 2.5, respectively.
|
|
SUVmax ≥ 2.0 (%)
|
SUVmax ≥ 2.5 (%)
|
Gender
|
Male
|
67 (71.3)
|
61 (69.3)
|
|
female
|
27 (28.7)
|
27 (30.7)
|
Age
|
≤ 65y
|
48 (51.1)
|
45 (51.1)
|
|
> 65y
|
46 (48.9)
|
43 (48.9)
|
Pathology
|
Squamous carcinoma
|
36 (38.3)
|
34 (38.6)
|
|
Adenocarcinoma
|
50 (53.2)
|
47 (53.4)
|
|
Small cell cancer
|
8 (8.5)
|
7 (8.0)
|
TNM stage
|
I-II
|
13 (13.8)
|
10 (11.4)
|
|
III-IV
|
81 (86.2)
|
78 (88.6)
|
Treatment
|
No treat
|
17 (18.1)
|
15 (17.0)
|
modality
|
Surgery only
|
9 (9.6)
|
7 (8.0)
|
|
Chemotherapy only
|
18 (19.1)
|
18 (20.5)
|
|
Target therapy only
|
7 (7.4%)
|
6 (6.8)
|
|
Combined therapy
|
43 (45.7)
|
42 (47.7)
|
Surgery
|
Yes
|
20 (21.3)
|
18 (20.5)
|
|
No
|
74 (78.7)
|
70 (79.5)
|
Radiation
|
Yes
|
33 (35.1)
|
32 (36.4)
|
therapy
|
No
|
61 (64.9)
|
56 (63.6)
|
Chemotherapy
|
Yes
|
59 (62.8)
|
58 (65.9)
|
|
No
|
35 (37.2)
|
30 (34.1)
|
Target therapy
|
Yes
|
20 (21.3)
|
18 (20.5)
|
|
No
|
74 (78.7)
|
70 (79.5)
|
Body weight
|
Mean ± SD
|
64.5 ± 10.3
|
64.3 ± 10.4
|
(kg)
|
Median
|
65.0
|
65.0
|
|
Range
|
36–90
|
36.0–90.0
|
Number of
|
Mean ± SD
|
5.2 ± 2.6
|
4.5 ± 2.6
|
foci
|
Median
|
4
|
3.5
|
|
Range
|
2–16
|
2–15
|
Volume of
|
Mean ± SD
|
41.4 ± 66.5
|
42.2 ± 68.0
|
Primary tumor
|
Median
|
20.0
|
25.0
|
(ml)
|
Range
|
0.8-548.9
|
0.8-548.9
|
SUVmax of
|
Mean ± SD
|
11.8 ± 5.8
|
11.8 ± 5.6
|
primary
|
Median
|
11.2
|
11.0
|
|
Range
|
2.6–27.0
|
2.6–27.0
|
TLG
|
Mean ± SD
|
368.2 ± 550.0
|
374.0 ± 572.5
|
|
Median
|
175.9
|
175.9
|
|
Range
|
15.5-4270.9
|
15.5-4270.9
|
Treatment Modalities
In the cohort 1, after PET/CT examination, there were 8, 5, 55 and 26 patients in TNM stage I to IV, successively. Among the patients in stage I-II, 9/13 accepted surgery only, and others received chemotherapy only (n = 1), target therapy only (n = 1), combined treatment (n = 1), and no treatment (n = 1). Correspondingly, 3/81 patients in stage III-IV underwent surgery only, and others received combined therapy. The treatment modalities (Fig. 2) did not exist significant difference between stage I and II (x2 = 0.612, p = 0.450) or stage III and IV (x2 = 8.482, p = 0.075), but did among stage I to IV (x2 = 23.759, p < 0.001). Therefore, in our analysis, TNM stages were stratified into I-II and III-IV. Additionally, none of this cohort was administrated radiotherapy alone which might result from the inclusion criteria that treatment after the PET/CT examination were followed.
Survival Analysis of Patient Characteristics
During a median observation time of 40 months (ranged from 34 to 55 months), both 70 patients of cohort 1 (n = 94) and 66 patients of cohort 2 (n = 88) died in a median OS of 10 months (ranged from 0 to 40 months). In the univariate Cox analysis, patient characteristics were separately regressed against OS. In general, the squamous and adenocarcinoma patients in early-stage who accepted surgery or target therapy had longer OS than others. On Fig. 3, hazard ratios (HR) of the factors are similar between the cohorts. Compared to no treatment, surgery and target therapy were significantly against OS in both cohorts. Additionally, although total lesion glycolysis (TLG) was a significant variable in both cohorts, its range of HR was very closed to 1.
In the multivariate Cox analysis of the two cohorts (Table 2), 3 patient characteristics (TNM stage, surgery and target therapy) were significantly against OS in both cohorts, but others were failed. Therefore, in further analysis, the factors of surgery, target therapy, and TNM stage were used as the basic model for the evaluation of the prognostic ability of FeatureSD.. The Harrell's concordance index (C-index) of the model composed of the 3 variables was 0.717 (95% CI: 0.662–0.772) and 0.723 (95% CI: 0.672–0.774) in the cohort 1 and the cohort 2, respectively. Above all, the criteria of malignant LN could slightly affect the results of survival analysis, and C-index of the regressed model were in an overlapped 95% CI.
Table 2
Multivariate Cox analysis of patient characteristics.
|
|
P value
|
HR (95% CI)
|
Cohort 1
|
Surgery
|
0.001
|
0.217 (0.090–0.528)
|
|
Target therapy
|
0.009
|
0.444 (0.241–0.818)
|
|
TNM stage
|
0.050
|
3.318 (0.978–11.254)
|
Cohort 2
|
Surgery
|
0.002
|
0.249 (0.105–0.591)
|
|
Target therapy
|
0.006
|
0.397 (0.206–0.765)
|
|
TNM stage
|
0.023
|
5.308 (1.256–22.427)
|
Feature Selection For The Combined And Thin-section Ct
On both the combined and thin-section CT images of the enrolled patients, 141 features (1683 variables) were extracted from each lesion, and, using the R package, SD of each feature among malignant foci for every individual (FeatureSD) was calculated as continuous variables. Subsequently, the survival extreme gradient boosting method (survival XGBoost) was used to score variables including continuous ones, and to rank the prognostic ability of each feature by the variable importance (VIMP).
In the survival XGBoost analysis of the data extracted from the combined CT images, the independent variables included FeatureSD and the 3 significant characteristics in the Cox regressions. On Figs. 4, 19 and 15 features with a predictive VIMP are selected for the cohort 1 and the cohort 2, respectively. Among the ranked variables, three were repetitive between the cohorts, and the measured correlation of GLCM (Gray Level Cooccurrence Matrix) ranked in the top three in both cohorts. However, none of the 3 patient characteristics were ranked high enough on the list. Above all, as continuous variables, some FeatureSDs among malignant foci were strong prognostic factors for lung cancer patients, and even had much higher VIMP than those identified by the Cox analysis.
By contrast, the data extracted from thin-section CT images indicated that, besides the 3 patient characteristics, no FeatureSD was repetitive between the top 20 VIMP lists of the cohort 1 and cohort 2. Detailed results were in the Supplementary Data (Fig. 1S). Therefore, only the data from the combined CT images was used in our further analysis.
Feature Evaluation
The FeatureSD identified in the survival XGBoost analysis ranged from 0 to 0.437 with a median of 0.163 (0.167 ± 0.116) and from 0 to 0.419 with a median of 0.141 (0.152 ± 0.114) in the cohort1 and 2, respectively. Furthermore, we used the penalized smoothing splines method to determine the threshold of the FeatureSD. Figure S2 indicates that, especially in the cohort 2, exponential HR (hazard ratio) increases with increase of the variable, and the reference value is 0.166 and 0.138 for the cohort 1 and 2, respectively. However, when using the reference values to group the patients, C-index of the feature could be even worse than directly using the continuous variable.
In the evaluation, the prognostic ability of the FeatureSD was compared to the 3 significant factors selected by the Cox regressions. On Fig. 5, C-index of the FeatureSD in the cohort 2 is the highest among the variables, and is the 2nd highest in the cohort 1. Furthermore, C-index of the model included FeatureSD was higher than others. Nomogram and calibration plots of the model integrated the FeatureSD (Fig. 6) indicated the reliability of the model. Above all, some FeatureSD (correlation of GLCM in this study) among malignant foci within an individual were powerful prognostic factors, and their prognostic ability was obviously higher than the factors of patient characteristics.