Association of immune-pathology features with survival

Resected tumor samples were stained for two pathology markers and the following metrics extracted: CD3 count (to assess lymphocyte infiltration) and % of tumor cells expressing PDL1 (to assess tumor-mediated immune suppression). In the training cohort, optimal cutoffs for both metrics were determined to maximize association with OS when partitioning this cohort into 4 immune pathology groups: PDL1 cutoff of 2% and a CD3 count cutoff of 1910cells/mm² (Fig. 1A). Representative IHC stains for CD3 and PDL1 from the 4 immune-pathology groups are presented in Fig. 1B. When considered separately, stratification by PDL1 or CD3 utilizing the identified cutpoints exhibited a trend towards a significant association with survival (P=0.08 and P=0.06, respectively). However, when considered together, pathology classification using both PDL1 and CD3 exhibited a significant association with survival; specifically, patients with immune-activated tumors (CD3^hiPDL1^lo) had the highest OS rate at 5yr (80%), and patients with immune-inhibited tumors (CD3^loPDL1^hi) had the worst OS rate at 5yr (35%) (P=0.04) (Fig. 1A). This association held on multivariate analysis considering baseline tumor and treatment factors (Supplemental Table ^S1).

Open in a separate window

Figure 1

Pathology grouping based on CD3+ infiltrating lymphocyte count and PD-L1+ tumor cell count in resected tumors showing pathology-based stratification and corresponding survival in the training cohort (A). Representative images showing immunohistochemistry PD-L1 and CD3 staining from patients within the training cohort (B). Pathology groupings and survival stratification are reproduced in the validation cohort (C).

Patients in the validation group were similarly stratified based on tumor CD3 and PDL1 staining (Fig. 1B,C). When considered separately, stratification by CD3 count exhibited a significant association with survival (P<0.01), while stratification by PDL1 did not (P=0.17). When considering both CD3 and PDL1 together, a similar association was noted between immune-pathology characteristics and survival; specifically, patients with immune-activated tumors had the highest 5-yr OS rate (95%), and the patients with immune-inhibited tumors had the lowest (54%) (P=0.002) (Fig. 1D). This association held on multivariate analysis (Supplemental Table ^S1). Concordance indices based on the immune-pathology subtypes alone were 0.60 for the training set and 0.63 for the validation set.

Creating the radiomics models

Hierarchical clustering of 12 prespecified radiomics features (see Methods) generated 8 independent radiomics models that were associated with immune-pathology features and/or OS (Fig. 2). Each model was produced by organizing all patients within the training dataset into radiomics clusters based on 4–5 radiomics parameters, which are a subset of the considered 12 parameters. For example, column 1 refers to radiomics model #1 which groups patients into clusters utilizing the voxel intensity metrics mean, kurtosis, and uniformity in addition to the gray level co-occurrence metrics of contrast and modified homogeneity. Of these 8 models, 3 were associated with both OS (P<0.05) and immune-pathology grouping (P<0.05), 1 of which was entirely composed of highly reproducible features (Supplementary Fig. ^S1). This model, which was associated with survival, immune pathology, and composed of highly reproducible features, is named the immune-pathology informed model (IPIM), was used for the remainder of the analyses.

Open in a separate window

Figure 2

Generation of 8 distinct radiomics models, each indicated by a column. Each model utilizes a subgroup of 4–5 radiomics parameter from the 12 considered radiomics features. When considering radiomics model, (*) indicates models that exhibit a significant association with overall survival (models 1–5) and (†) indicates models that exhibit a significant association with immune pathology (models 1–3 and 6–8). When considering radiomics feature, (‡) indicates features that are highly reproducible with <3% inter-reader variability (voxel intensity mean, entropy, standard deviation, and uniformity and gray level co-occurrence homogeneity and modified homogeneity).

The IPIM consists of 4 features, 3 of which are derived by first-order analysis of image-feature intensity: mean, standard deviation, and uniformity (Fig. 3). The fourth feature is a second-order analysis of the homogeneity of the derived gray level co-occurrence matrix. IPIM features were organized into 4 clusters, named A through D. A web application was created to assign an IPIM designation for any image when the value of these 4 features was input to the application (URL: https:/biostatistics.mdanderson.org/RadiomicsSubtypingCalculator).

Open in a separate window

Figure 3

Immune-pathology informed radiomics model creation and representative radiology images (dotted bar=1cm) in the training cohort; green and red represent high and low value based on feature Z-score, respectively. The mean/median feature values are displayed for each radiomics cluster within the heatmap (A). IPIM cluster pathology characteristics shown as median (1^st and 3^rd quartiles) percent tumor PD-L1 expression and CD3 count (B). Probability of a patient being assigned to cluster D assignment given specified CD3 count and log (%PDL1 positivity). High and low probability designated by red and blue, respectively (C). Overall survival stratified by IPIM model assignment (D).

Multinomial regression for distinguishing the 4 IPIM clusters identified significant differences in PDL1 and CD3 expression (P=0.01). These pathology characteristics are described for each cluster. Significant differences in OS were also observed among IPIM clusters (P=0.01). Cluster D had a predominantly immune-activated phenotype, with high CD3 and low PDL1 expression (Fig. 3B–D), and 5-yr OS rates were correspondingly higher among patients grouped into Cluster D (91%) than for patients in Cluster A (61%), Cluster B (41%), and Cluster C (50%). This association of Cluster D with improved OS held on multivariate analysis of the training set (Table 2). The c-indices for the univariate and multivariate models were 0.61 and 0.70, respectively.

Table 2

Multivariate associations with survival.

Characteristic	Training Set (n=114)			Validation Set (n=176)			Validation Set (Stage 1 only; n=111)
	HR	95% CI	P Value	HR	95% CI	P Value	HR	95% CI	P Value
Lesion size (per cm)	1.19	1.04–1.36	0.01	1.06	0.92–1.21	0.43	1.13	0.76–1.69	0.54
N1/N2 (vs. N0)	1.36	0.77–2.40	0.30	3.75	2.07–6.81	<0.01	—	—	—
Adjuvant therapy (vs. none)	0.77	0.43–1.36	0.37	0.63	0.34–1.19	0.16	0.47	0.06–3.99	0.49
Age at surgery (per year)	1.04	1.01–1.07	0.02	1.02	0.99–1.05	0.13	1.06	1.01–1.12	0.03
Squamous histology	0.93	0.56–1.56	0.79	1.70	1.01–2.88	0.047	1.68	0.79–3.57	0.18
Current smoker (vs. former and never smoker)	1.37	0.80–2.30	0.26	1.26	0.75–2.10	0.38	0.78	0.34–1.81	0.57
Radiomics cluster (REF: Cluster D)
Cluster A	1.52	0.42–5.46	0.52	1.44	0.68–3.08	0.34	2.91	1.15–7.36	0.02
Cluster B	3.45	1.01–11.78	0.048	1.54	0.73–3.24	0.25	1.14	0.32–4.04	0.84
Cluster C	2.81	0.80–9.86	0.11	2.75	1.24–6.10	0.01	5.30	2.08–13.50	<0.01

Open in a separate window

Abbreviations: 95% CI, 95% confidence interval; HR, hazard ratio.

Pathology analysis

Four-micron-thick sequential histologic tumor sections were obtained from a representative formalin-fixed, paraffin-embedded tumor block and used for IHC analysis. IHC staining was done using an automated staining system (BOND-MAX; Leica Biosystems, Nussloch, GmbH) with antibodies against PDL1 (clone E1L3N, dilution 1:100; Cell Signaling Technology, Beverly, MA, USA) and CD3 (T-cell lymphocytes; dilution 1:100; Dako, Carpinteria, CA, USA). Marker expression was detected with a Novocastra Bond Polymer Refine Detection kit (Leica Biosystems) with a diaminobenzidine reaction to detect antibody labeling and hematoxylin counterstaining. The slides were then scanned in an Aperio AT2 scanner (Leica Biosystems). CD3+ cells were counted by a pathologist using Aperio Image Toolbox analysis software (Aperio, Leica Biosystems) and expressed as cell density (CD3+ cells/mm² of analyzed tissue). Percentage of tumor cells staining for PDL1 was quantified with the GENIE histology pattern recognition algorithm (Aperio) for automated identification of the tumor regions, under a pathologist’s supervision, via a membrane algorithm. Further details regarding staining technique are provided in the following references^32,33.

Patients were grouped based on PDL1 staining in tumor cells and characteristics of tumor-infiltrating lymphocytes (TILs) into 4 previously established pathology groups based on T-cell infiltration and tumor PDL1 expression as follows: PD-L1^hiCD3^hi, PD-L1^loTIL^lo, PD-L1^hiTIL^lo, PD-L1^loTIL^hi (Fig. 1)²⁵. Optimal immune subtypes were identified from the training set with thresholds of PDL1 and CD3 expression yielding maximum association with OS as characterized by the likelihood-ratio test based on the Cox proportional hazards model. Cutpoints identified in the training cohort were then applied to the validation cohort.

Radiomics feature extraction and characterization

Pretreatment chest CT scans with contrast agent were obtained at our institution and contoured by using 3D Slicer^34,35. Radiomics features were extracted from contoured tumors by using IBEX (example contours displayed in Supplemental Fig. ^S2)³⁶. We initially investigated the full set of 490 available parameters available via IBEX. Notably, radiomics analyses produce a large number of highly redundant feature sets, which necessitates variable selection to produce a sparse subset. This analytical approach, when applied to highly correlated features, often fails to produce reproducible findings across studies. Moreover, radiomics parameters are sensitive to filtering and preprocessing techniques. To overcome these limitations, we evaluated prognostic signatures on the basis of 12 features previously studied in NSCLC. These features are based on calculations from the voxel intensity: mean, kurtosis, skewness, entropy, standard deviation, and uniformity and on calculations from the gray level co-occurrence matrix: contrast, dissimilarity, energy, entropy, homogeneity, and modified homogeneity. Mathematic definitions for each feature along with the utilized filtering and preprocessing methodlogy are provided in Supplementary Tables ^S3–⁴ Features were assessed for their inter-user reproducibility by having 3 physicians use two different programs (3D Slicer and MIM Vista^34,35,38,39) to contour a random set of 10 patients. Features exhibiting <3% deviation attributable to inter-user heterogeneity in both programs were considered “highly reproducible” and thereby given preference for radiomics model selection. The resultant Bland-Altman plots are reported in Supplementary Fig. ^S1.

We gratefully acknowledge help from Christine Wogan in preparing this manuscript. We also acknowledge support from the Andrew Sabin Family Fellowship, Center for Radiation Oncology Research, the Sheikh Ahmed Center for Pancreatic Cancer Research, institutional funds from the The University of Texas MD Anderson Cancer Center, equipment support from GE Healthcare and the Center for Advanced Biomedical Imaging (CABI), Philips Healthcare, and Cancer Center Support (Core) Grant CA016672 from the National Cancer Institute to MD Anderson. Dr. Chad Tang was supported by a Young Investigator Award from the American Society of Clinical Oncology and Radioloigcal Society of North America. Dr. Eugene Koay was supported by NIH (U54CA210181-01, U54CA143837, U01CA200468, and U01CA196403), the Pancreatic Cancer Action Network (14-20-25-KOAY and 16-65-SING), and the Radiological Society of North America (RSD1429).