In modern oncology, a patient’s treatment plan is dictated by their tumor’s genetic "fingerprint." Knowing whether a glioma has an IDH mutation or if a lung cancer expresses EGFR determines which drugs will work and how long a patient might survive. Traditionally, the only way to get this information is through a physical biopsy—drilling into the skull or inserting needles into organs to extract tissue.
However, a growing body of research suggests that we can perform "virtual biopsies" by decoding the high-dimensional data already hidden within standard medical images like MRI, CT, and PET scans. This field, known as radiogenomics, treats medical images not just as pictures, but as spatially-resolved data maps of a tumor’s underlying biology.
The Biological Bridge: From Genotype to Phenotype
The core intuition behind radiogenomics is that a microscopic genetic mutation does not exist in a vacuum; it fundamentally alters the tumor's macroscopic environment. These alterations—changes in blood flow, cell density, and metabolic waste—manifest as subtle patterns in medical imaging.
A clear example of this mechanism is found in the study of gliomas. Philipp Kickingereder and Wolfgang Wick demonstrated in IDH mutation status is associated with a distinct hypoxia/angiogenesis transcriptome signature that IDH mutations act as a biological "brake" on angiogenesis (the growth of new blood vessels). Because IDH-mutant tumors grow fewer vessels, they have lower blood volume. By measuring the relative cerebral blood volume (rCBV) via non-invasive MRI, researchers can predict the mutation status with high accuracy because the imaging is capturing a direct physical consequence of the genetic code.
Decoding the "Radiomic" Signature
While some signs are visible to the naked eye, most radiogenomic information is "sub-visual." Researchers use radiomics—the high-throughput extraction of thousands of quantitative features—to identify textures and shapes that human radiologists might miss.
Paul Eichinger and colleagues showed in Diffusion tensor image features predict IDH genotype in newly diagnosed WHO grade II/III gliomas that machine learning models can analyze complex texture patterns in Diffusion Tensor Imaging (DTI) to predict genotypes with up to 95% accuracy. Similarly, Xiao Feng utilized multiregional MRI analysis in Preoperative prediction of MGMT promoter methylation in glioblastoma to show that the relationship between different tumor subregions (the core vs. the edema) provides critical clues about MGMT methylation, a key marker for chemotherapy response.
Virtual Biopsy Maps and Spatial Heterogeneity
One major limitation of a physical biopsy is that it only samples a tiny fraction of the tumor. If a tumor is heterogeneous—meaning different parts have different mutations—a single needle strike might miss the most aggressive cells.
Radiogenomics solves this by providing a "global" view. Marwa Ismail and Ramon Correa developed the SpACe (Spatial-And-Context aware) framework in Spatial-And-Context aware "virtual biopsy" radiogenomic maps. This approach creates a "virtual biopsy map" across the entire tumor volume, combining local textures with population-level spatial data to pinpoint exactly where specific mutations are likely occurring.
Multi-Modal Fusion
The most powerful models don't rely on a single type of scan. Just as a doctor uses multiple senses to diagnose a patient, researchers are using "multi-modal" deep learning to combine different imaging "senses."
In Multi-modal learning for predicting the genotype of glioma, Yiran Wei and Lei Zhu demonstrated that combining 3D structural MRI with geometric "point clouds" of tumor shape and global brain connectivity networks significantly outperforms models using images alone. This holistic view captures how the tumor is not just an isolated mass, but a biological entity interacting with the entire organ.
Expanding the Reach of Virtual Biopsies
While neuro-oncology has been the primary testing ground, the "virtual biopsy" concept is expanding rapidly to other cancers:
- Lung Cancer: Lilan She and Yunjing Xue used spectral CT and AI to predict EGFR mutations in lung adenocarcinoma (Predicting EGFR gene mutation in lung adenocarcinoma).
- Liver Cancer: Fengxia Zeng explored how radiomics can predict $\beta$-catenin mutations in hepatocellular carcinoma to help select patients for immunotherapy (Preoperative radiomics model using gadobenate dimeglumine-enhanced MRI).
- Breast Cancer: Chunping Mao utilized Diffusion Spectrum Imaging to non-invasively assess HER2 status (Quantitative Parameters of Diffusion Spectrum Imaging).
The Future: Quantitative and Sex-Specific Models
As the field matures, researchers are moving toward more robust, standardized methods. Manabu Kinoshita argues in Advances in the Qualitative Diagnosis of Glioma that the future lies in quantitative MRI, which measures absolute physical properties (like T1 and T2 relaxation times) rather than relative signal intensities. This would allow models to work across different hospitals and scanners.
Furthermore, László Papp has highlighted that radiomic signatures can differ significantly between men and women (Sex-specific radiomic features of PET in patients with newly-diagnosed gliomas), suggesting that the next generation of "virtual biopsies" will need to be as personalized as the genetic therapies they aim to guide.