Traditional radiology operates on a "perceptual ceiling." When a radiologist looks at an MRI or CT scan, they are trained to identify structural anomalies—the size of a tumor, the presence of edema, or the blockage of a vessel. However, as Eros Montin and colleagues point out in Editorial: Radiomics and AI for clinical and translational medicine, these images contain high-dimensional quantitative data—complex textures and spatial pixel distributions—that are entirely invisible to the human eye.
The research direction of Radiomics seeks to break through this ceiling. By treating medical images not as pictures, but as data, researchers are extracting "latent" features to predict a patient’s molecular status, treatment response, and survival without the need for invasive biopsies.
The Intuition: Images as Data Reservoirs
To understand why radiomics works, we must view a pixel not just as a shade of gray, but as a data point in a larger mathematical landscape. A tumor that looks uniform to a human may actually possess subtle "heterogeneity"—minute variations in intensity and texture that reflect its underlying biology.
For instance, Cecilia S. Lee and Ariel S. Rokem demonstrated that deep learning can "infer" functional blood flow from static structural images of the eye Generating retinal flow maps from structural optical coherence tomography with artificial intelligence. The algorithm detects subtle speckle patterns indicating physiological movement that a human expert simply cannot perceive. This principle applies across modalities: whether it is ultrasound for thyroid cancer Fu Li et al., Using ultrasound features and radiomics analysis to predict lymph node metastasis in patients with thyroid cancer or CT scans for lung cancer Chen Huang et al., A Computer Tomography Radiomics-based model to Predict Survival of Patients with EGFR‑Mutated Non‑small‑cell Lung Cancer, the core mechanism is the same: identifying sub-visual mathematical regularities that correlate with biological reality.
Non-Invasive Molecular Profiling
One of the most powerful applications of this field is the "virtual biopsy." In brain cancer, knowing the genetic makeup of a tumor (like the IDH genotype or MGMT methylation status) is critical for treatment, but usually requires risky surgery.
Shuang Wu and Jin Meng showed that machine learning models, specifically Random Forest, can predict IDH genotypes by analyzing these hidden textural patterns Radiomics-based machine learning methods for isocitrate dehydrogenase genotype prediction of diffuse gliomas. Similarly, Xiao Feng used multi-sequence MRI to predict MGMT promoter methylation, effectively decoding the tumor's molecular signature from standard clinical scans Preoperative prediction of MGMT promoter methylation in glioblastoma based on multiregional and multi-sequence MRI radiomics analysis.
Predicting the Future: Prognosis and Response
Beyond diagnosis, radiomics offers a window into a patient's future. By extracting "deep radiomic signatures," researchers like Ahmad Chaddad have linked imaging patterns to immune cell markers to predict survival outcomes in glioma patients Deep radiomic signature with immune cell markers predicts the survival of glioma patients.
This predictive power extends to treatment response. Fang Wang and Hong Yang developed a model using pre-treatment CT scans to predict how lung cancer patients would respond to chemoimmunotherapy A combined model using pre-treatment CT radiomics and clinicopathological features of non-small cell lung cancer to predict major pathological responses after neoadjuvant chemoimmunotherapy. The intuition here is that the "spatial heterogeneity"—how messy or organized a tumor looks at a pixel level—indicates how aggressive the cancer is and how likely it is to succumb to specific drugs.
Opening the Black Box
A major hurdle in this field is the "black box" nature of deep learning. If a model predicts a high risk of death, a clinician needs to know why. Tobias Weber and David Rügamer addressed this with CoxVAE, a framework that generates survival-oriented embeddings Survival-oriented embeddings for improving accessibility to complex data structures. This allows researchers to visualize what the model considers a "high-risk" image, moving from pure prediction toward interpretable clinical tools.
However, the field remains grounded by healthy skepticism. Laurens S. ter Maat and colleagues found that while deep learning can find hidden patterns, it does not always outperform simple clinical data like blood tests Deep learning on CT scans to predict checkpoint inhibitor treatment outcomes in advanced melanoma. This suggests that the future of the field lies in hybrid models—combining the sub-visual insights of radiomics with the established wisdom of clinical practice.