Using 190 TCGA-KIRC patients whose tumors we fully segmented on multiphase CT, we extracted 2,824 quantitative texture features, pared them to the 388 most scanner-reproducible metrics, and trained a random-forest model to predict the expression of thirteen hypoxia-linked genes previously associated with prognosis and treatment response.
Our findings were modest but encouraging. Texture patterns, especially second-order measures such as gray-level dependence (GLDM) and co-occurrence (GLCM), showed statistically significant correlations with several key biomarkers. The model best captured under-expression of the tumor-suppressor KLF6 (r = 0.27) and apoptosis regulator BCL2 (r = 0.19), as well as expression of the hypoxia-responsive transcription factor ETS1 (r = 0.25). Notably, these associations strengthened in high-grade tumors, where hypoxia biology is most pronounced; KLF6 prediction, for example, rose to r = 0.35 in Fuhrman grade 3/4 disease. Conversely, high PLOD2, an enzyme driving hypoxia-induced collagen cross-linking, is tracked with lower texture-uniformity signals, suggesting that structural remodeling leaves a distinct radiomic fingerprint.
This is significant because the most common current diagnostic modality, tissue biopsy, only samples a tiny slice of a genetically heterogeneous tumor. Liquid biopsies present a potential alternative, but large-scale genetic profiling remains cost-prohibitive for many centers.
Radiomics, in contrast, piggybacks on scans every ccRCC patient already receives, interrogating the whole lesion and its micro-environment at no extra patient risk. If further validated, a texture-based hypoxia score could refine risk stratification, flag patients who might benefit from HIF-targeted therapies or dose-escalated stereotactic radiotherapy, and even monitor biologic response over time without repeat biopsy.
The study does have caveats. Our cohort, though the largest publicly available with paired imaging and RNA-seq, is still single-institution and predominantly Caucasian; multi-center external validation is essential. Additionally, the cross-sectional nature of our data limits tracking over time; longitudinal imaging could address this. Finally, radiomic features are abstract; linking them to spatial transcriptomics or functional MRI could sharpen biological interpretability.
Still, the signal is clear: standard CT harbors more biologic insight than meets the eye. Texture analysis will not replace genomics, but it can meaningfully augment it, especially where resources or tissue are scarce. Future directions can include automating segmentation, expanding to perfusion MRI, and prospectively testing whether the radiomic hypoxia signature predicts hypoxia-related gene expression and clinical outcomes in nephrectomy and systemic-therapy cohorts. We look forward to seeing whether these early correlations can ultimately translate into clinically actionable stratification and improved care for patients with ccRCC.
Written by: Yijun Shao,1 Harmony S Cen,1,2 Anu Dhananjay,3 S J Pawan,4 Xiaomeng Lei,4 Inderbir S Gill,5 Anishka D'souza,6 Vinay A Duddalwar4,5,7,8
- Keck School of Medicine, University of Southern California, Los Angeles, CA, USA.
- Mark and Mary Stevens Neuroimaging and Informatics Institute, University of Southern California, Los Angeles, CA, USA.
- Newport High School, Bellevue, WA, USA.
- Radiomics Lab, Department of Radiology, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA.
- Institute of Urology, University of Southern California, Los Angeles, CA, USA.
- Department of Medical Oncology, Keck School of Medicine, University of Southern California, Los Angeles, CA, USA.
- Alfred E. Mann Department of Biomedical Engineering, University of Southern California, Los Angeles, CA, USA.
- Department of Radiology, Los Angeles General Medical Center, Los Angeles, CA, USA.