Radiogenomics

In short: Radiogenomics asks whether what we see on an image relates to what is happening in a tumour’s genes. The word also has a second, older meaning in radiation oncology.

Two meanings of one word

In medical imaging, radiogenomics (sometimes called imaging genomics) studies associations between imaging features and genomic or molecular features of disease. A typical question is whether a pattern on CT or MRI tends to appear more often in tumours that carry a particular mutation or expression profile.

In radiation oncology, the same word has long described something different: how inherited genetic variation relates to a person’s response to radiotherapy, especially side effects in normal tissue. When you read a paper, check which meaning the authors use.

Why people study it

Imaging is non-invasive, can cover a whole tumour and can be repeated over time. Tissue sampling looks at a small piece at one moment. Researchers hope imaging might one day add context about heterogeneity or change over time. That hope is a research direction, not an established clinical replacement for molecular testing.

What it is not

• Not a way to read a mutation from a scan in routine care.
• Not proof that an imaging feature causes, or is caused by, a genomic change.
• Not a substitute for biopsy, pathology or validated molecular tests.

Takeaways

• Radiogenomics studies associations between imaging and molecular features.
• Published associations vary in strength and reproducibility.

In short: An imaging phenotype is what a tumour looks like on a scan, described by a reader or measured by software. A genomic phenotype is its molecular make-up.

Semantic features: described by people

Radiologists describe lesions with words such as margin, enhancement pattern, necrosis or oedema. When these descriptions are standardised into a fixed vocabulary they become semantic features that can be analysed. In glioma research, for example, a published feature set called VASARI was designed so that different readers score MRI appearances the same way.

Quantitative features: measured by software

Software can compute size, shape, intensity statistics and texture from a segmented region. These are quantitative or radiomic features. They are objective in the sense that the same software on the same data gives the same number, but they are sensitive to how the image was acquired and processed.

Genomic and molecular phenotypes

On the molecular side, phenotypes include mutations, copy-number changes, gene expression, methylation and protein expression. Each is measured by a laboratory method with its own error and its own sampling limits: a biopsy captures one region of a tumour that may be heterogeneous.

Takeaways

• Imaging phenotypes can be semantic (reader-described) or quantitative (software-measured).

In short: Radiomics extracts many quantitative features from images. Radiogenomics relates imaging features, radiomic or semantic, to molecular data.

Radiomics

Radiomics turns images into large tables of numbers describing shape, intensity and texture. Those numbers can be related to many outcomes: diagnosis, survival, response, or molecular status.

Radiogenomics

Radiogenomics is one use of imaging features: relating them to genomic or molecular data. It can use radiomic features, semantic features, or both. Not all radiomics is radiogenomics, and radiogenomics does not require radiomics.

Why the distinction matters when you search

A paper that predicts survival from radiomic features is radiomics but not radiogenomics. A paper relating a reader-scored MRI sign to IDH status is radiogenomics without high-dimensional radiomics. RadioGx labels both and does not merge them.

In short: A tour of genes that appear often in radiogenomics research, with the caution each one needs.

Lung adenocarcinoma: EGFR and KRAS

Studies have looked for CT features associated with EGFR or KRAS mutations, such as the proportion of ground-glass opacity or particular shapes. Results differ between cohorts and populations, and none replaces molecular testing, which remains essential for treatment decisions.

Glioma: IDH and 1p/19q

In diffuse gliomas, an MRI appearance called the T2-FLAIR mismatch sign has been reported in several studies as highly specific for a subgroup of IDH-mutant tumours without 1p/19q co-deletion, but it is present in only a minority of them. Specific is not the same as sensitive: its absence tells you little.

Glioblastoma: MGMT promoter methylation

Many studies have tried to predict MGMT promoter methylation from MRI. Reported performance varies widely and external validation is inconsistent. Methylation is established by laboratory testing of tissue.

Kidney cancer: VHL, BAP1 and others

Public datasets that link TCGA kidney cases to imaging have supported exploratory work on features associated with mutations such as BAP1. These are hypothesis-generating findings.

Takeaways

• Imaging-genomic associations for these genes are reported in research literature with variable reproducibility.

In short: Why a significant p-value in one dataset is the start of a story, not the end, and how RadioGx shows that difference.

From association to tool

A finding moves through stages: discovery in one cohort, confirmation in an independent cohort, prospective testing, and finally evaluation of whether using it improves decisions or outcomes. Most imaging-genomic findings sit at the first or second stage.

Common traps

• Overfitting: many features, few patients, a model that memorises noise.
• Leakage: test data influencing feature selection or tuning.
• Multiple comparisons: testing hundreds of features and reporting the winners.
• Publication bias: positive findings are more likely to be published.
• Spectrum bias: cohorts that are cleaner than real practice.

How RadioGx labels evidence

RadioGx marks every relationship as source-reported, reference index, derived or education context, and never upgrades a "possible" match to "exact". Interpretation scores describe how RadioGx read your query, never how true a clinical claim is.