The Diversity Hidden Within Stage T1 Lung Cancer Clinical T1 stage lung adenocarcinoma (LUAD) - tumors up to 3 cm - is often considered a single category for treatment planning. However, within this seemingly uniform group, there is wide biological variation: some tumors are indolent and nearly curable by surgery alone, while others harbor aggressive pathological features like spread through airspaces (STAS), lymph node metastasis (LNM), and visceral pleural invasion (VPI). Standard clinical staging cannot reliably distinguish these subgroups.
Radiomics as a Window into Tumor Biology Radiomic features extracted from CT images quantify texture, shape, and intensity characteristics of tumors that are invisible to the naked eye but correlate with underlying histological and molecular properties. By clustering patients based on their radiomic profiles, it may be possible to identify biologically distinct subtypes that predict different clinical outcomes and require different treatment strategies.
Consensus Clustering Approach Unlike supervised machine learning (which requires labeled outcome data for training), consensus clustering is an unsupervised method that identifies natural groupings within the data. It repeatedly resamples the dataset and builds clusters each time, then identifies which patient pairs consistently cluster together. Groups that emerge robustly across many resampling iterations represent true biological subpopulations rather than statistical artifacts.
Study Population This retrospective study enrolled 497 patients with clinical T1 stage LUAD who underwent surgical resection. CT radiomic features were extracted from preoperative scans and consensus clustering was applied to identify optimal patient groupings, which were then correlated with pathological findings and molecular data.
Radiomic Feature Extraction From each patient's preoperative CT scan, hundreds of radiomic features were extracted encompassing first-order statistics (capturing intensity distribution), shape features (capturing 3D geometry), and texture features from matrices like GLCM and GLRLM (capturing spatial relationships between pixel intensities). Features were standardized and highly correlated features were filtered to reduce redundancy.
Consensus Clustering Protocol The consensus clustering algorithm was run across a range of cluster numbers (k = 2 to 10) using repeated subsampling. The optimal number of clusters was determined by examining the consensus cumulative distribution function - looking for a plateau that indicates stable, well-separated groups. The analysis identified k = 2 as the optimal solution, meaning patients naturally divided into two radiomic subgroups.
Pathological Correlation After clustering, the proportions of aggressive pathological features were compared between Cluster 1 and Cluster 2. Features examined included predominant histological growth pattern (solid, micropapillary versus lepidic, acinar), Ki-67 proliferation index, STAS, LNM, and VPI. Significant enrichment of any of these features in one cluster would validate the biological meaning of the radiomic separation.
Molecular Profiling Next-generation sequencing data was available for a subset of patients, allowing comparison of driver mutation frequencies (EGFR, KRAS, ALK, HER2, and others) between clusters. Differences in mutation frequencies would indicate that radiomic clusters capture distinct molecular landscapes rather than merely imaging noise.
Two Distinct Radiomic Subtypes Consensus clustering robustly identified k = 2 as the optimal solution. Cluster 1 contained tumors with higher radiomic heterogeneity, higher attenuation values, and denser texture signatures. Cluster 2 contained tumors with more homogeneous, ground-glass-like radiomic profiles. This imaging dichotomy turned out to correspond to meaningful biological differences.
Pathological Aggressiveness in Cluster 1 Cluster 1 tumors showed significantly higher rates of aggressive pathological subtypes - specifically micropapillary and solid predominant growth patterns. Elevated Ki-67 indices (indicating rapid proliferation), higher rates of STAS, LNM, and VPI were all significantly enriched in Cluster 1 compared to Cluster 2. This constellation of features defines a high-risk pathological phenotype linked to recurrence and metastasis.
HER2 Mutation as the Molecular Discriminator Among all driver mutations analyzed, HER2 mutation was the only one with significantly different frequency between the two clusters (p < 0.001), with HER2-mutant tumors preferentially clustering in the aggressive Cluster 1 group. This finding is biologically plausible given HER2's role in promoting proliferation and invasion, and has potential therapeutic implications as HER2-targeted agents become available for lung cancer.
EGFR and Other Mutations Despite EGFR being the most common driver mutation in lung adenocarcinoma, EGFR mutation frequency did not significantly differ between clusters. This suggests that radiomic clustering captures biological heterogeneity along axes other than the dominant driver mutation, reflecting aspects of tumor microenvironment, growth pattern, and stromal architecture.
Non-Invasive Risk Assessment The most important clinical implication of this work is that radiomic cluster assignment can potentially be determined from preoperative CT alone - without requiring biopsy or intraoperative assessment. Identifying Cluster 1 patients before surgery could inform the extent of resection (sublobar versus lobectomy), lymph node sampling strategy, and whether adjuvant therapy should be planned upfront.
Surgical Planning Applications Current surgical decision-making for T1 LUAD weighs risks and benefits of sub-lobar versus lobar resection based largely on tumor size. Radiomic cluster assignment adds a biology layer to this decision: Cluster 1 patients with high-risk imaging phenotype might benefit from more extensive lymph node dissection even when tumor size alone would not mandate it.
Identifying Candidates for Adjuvant Therapy T1 stage LUAD is typically considered low-risk, with adjuvant chemotherapy reserved for more advanced stages. However, Cluster 1 patients - particularly those with LNM or STAS confirmed pathologically - might benefit from consideration of adjuvant osimertinib (in EGFR-mutated cases) or immune checkpoint inhibitors based on their aggressive biological profile.
HER2 Targeting Opportunity The association between HER2 mutation and the aggressive radiomic cluster raises the possibility of preoperatively identifying HER2-mutant patients for enrollment in clinical trials of HER2-targeted therapies (trastuzumab deruxtecan has shown activity in HER2-mutant NSCLC). Radiomic pre-screening could supplement or prioritize molecular testing.
Retrospective Design Limitations This study was retrospective and single-institution, meaning selection bias and institutional practice patterns may influence findings. Patients who had surgical resection represent a more favorable subset of T1 LUAD, and results may not apply to the substantial proportion managed with stereotactic body radiation therapy or active surveillance.
External Validation Priority Consensus clustering results are known to be dataset-dependent. The two-cluster solution identified in this Chinese cohort must be validated in independent datasets from different geographic and ethnic populations to confirm that the radiomic clusters represent true biological archetypes rather than population-specific patterns.
Survival Data Needed While the study established that Cluster 1 has aggressive pathological features, direct survival outcome data (disease-free survival, overall survival) were not the primary endpoint. Subsequent analysis should confirm that Cluster 1 patients actually have worse outcomes after surgery and identify whether cluster-based risk stratification adds independent prognostic information beyond stage and pathology.
Integration with Multi-Omics Future studies should combine radiomic clustering with transcriptomic, proteomic, and spatial transcriptomics data to build comprehensive tumor phenotype maps. Understanding why certain radiomic textures correlate with STAS or HER2 mutation will require mechanistic studies connecting imaging physics with molecular biology.