CT-based Radiogenomic Analysis of Clinical Stage I Lung Adenocarcinoma with Histopathologic Features and Oncologic Outcomes

Abstract

Background A preoperative predictive model is needed that can be used to identify patients with lung adenocarcinoma (LUAD) who have a higher risk of recurrence or metastasis. Purpose To investigate associations between CT-based radiomic consensus clustering of stage I LUAD and clinical-pathologic features, genomic data, and patient outcomes. Materials and Methods Patients who underwent complete surgical resection for LUAD from April 2014 to December 2017 with preoperative CT and next-generation sequencing data were retrospectively identified. Comprehensive radiomic analysis was performed on preoperative CT images; tumors were classified as solid, ground glass, or mixed. Patients were clustered into groups based on their radiomics features using consensus clustering, and clusters were compared with tumor genomic alterations, histopathologic features, and recurrence-specific survival (Kruskal-Wallis test for continuous data, χ² or Fisher exact test for categorical data, and log-rank test for recurrence-specific survival). Cluster analysis was performed on the entire cohort and on the solid, ground-glass, and mixed lesion subgroups. Results In total, 219 patients were included in the study (median age, 68 years; interquartile range, 63-74 years; 150 [68%] women). Four radiomic clusters were identified. Cluster 1 was associated with lepidic, acinar, and papillary subtypes (76 of 90 [84%]); clusters 2 (13 of 50 [26%]) and 4 (13 of 45 [29%]) were associated with solid and micropapillary subtypes (P < .001). The EGFR alterations were highest in cluster 1 (38 of 90 [42%], P = .004). Clusters 2, 3, and 4 were associated with lymphovascular invasion (19 of 50 [38%], 14 of 34 [41%], and 28 of 45 [62%], respectively; P < .001) and tumor spread through air spaces (32 of 50 [64%], 21 of 34 [62%], and 31 of 45 [69%], respectively; P < .001). STK11 alterations (14 of 45 [31%]; P = .006), phosphoinositide 3-kinase pathway alterations (22 of 45 [49%], P < .001), and risk of recurrence (log-rank P < .001) were highest in cluster 4. Conclusion CT-based radiomic consensus clustering enabled identification of associations between radiomic features and clinicalpathologic and genomic features and outcomes in patients with clinical stage I lung adenocarcinoma. © RSNA, 2022 Online supplemental material is available for this article. See also the editorial by Nishino in this issue.

Graphical abstract

Figure 1:

Conditional density function demonstrating consensus clustering of radiomic features at a k value of 4. For a set of possible k values of 1–6, we iteratively subsampled 80% of the data set, hierarchically clustered each subsample, and then assessed the relative frequency that each patient was clustered with each other patient for each k value, resulting in a consensus matrix. For each k value, we then computed the change in area under the cumulative distribution function curve of the consensus matrix distribution, which showed how well separated the clusters were. We chose an optimal k value that corresponded to a sharp decrease in this change in area under the receiver operating characteristic curve, which indicated that further gains in separability were negligible after that k value.

Figure 2:

Consort diagram. IMPACT = Integrated Mutation Profiling of Actionable Cancer Targets.

Figure 3:

OncoPrint of clinical-pathologic and genomic variables for all patients with clinical stage I lung adenocarcinoma broken down by radiomic consensus clustering. Columns represent patients within each cluster of radiomics features, and rows represent the frequencies of alterations of all individually analyzed oncogenic genes and 10 canonical oncogenic pathways. Cluster characteristics were compared using Fisher or χ² exact test for categorical data and Kruskal-Wallis test for continuous data. * Associations between radiomic clusters and clinicopathologic or genomic variables are significant (P < .05). FGA = fraction of genome altered, LVI = lymphovascular invasion, NA = not applicable, MIP = micropapillary, SOL = solid, STAS = spread through air spaces, TMB = tumor mutational burden.

Figure 4:

Thin-section CT images of lesions used for radiomic cluster analysis. (A) Cluster 1. Ground-glass nodule in the right lower lobe (arrow) measuring 2.2 cm, predominantly lepidic histologic subtype, EGFR, and phosphoinositide 3-kinase (pPI3K) positive. Follow-up was 35 months without evidence of recurrence. (B) Cluster 2. Part-solid nodule in the left upper lobe (arrow) measuring 2.1 cm, solid and micropapillary histologic subtype, spread through air spaces (STAS) positive, STK11 and KRAS positive. Follow-up was 40 months without evidence of recurrence. (C) Cluster 3. Solid nodule (arrow) with thin ground-glass halo in the right lower lobe measuring 1.0 cm; solid and micropapillary histologic subtype; STAS positive; TP53, pPI3K, and receptor tyrosine kinase-Ras positive. The patient underwent right lower lobe wedge resection and was diagnosed with recurrence 9 months after resection. (D) Cluster 4. Solid nodule (arrow) in the right upper lobe measuring 1.4 cm; solid and micropapillary histologic subtype; STAS and lymphovascular invasion positive; STK11, KRAS, and pPI3K positive. The patient underwent right upper lobectomy and was diagnosed with recurrence 15 months after resection.

Figure 5:

Two-year recurrence-specific survival for each of the four radiomic consensus clusters of the overall study cohort.