Pancreatic ductal adenocarcinoma (PDAC) is one of the most lethal malignancies, with a five-year survival rate below 12%. A major reason for this poor prognosis is that most patients are diagnosed at locally advanced or metastatic stages, when curative surgery is no longer possible.
Contrast-enhanced computed tomography (CE-CT) is the primary imaging modality used for staging, resectability assessment, and treatment planning in PDAC. However, standard radiological reading relies on subjective visual interpretation and may not fully exploit the quantitative information encoded in the raw pixel data of CT images.
Radiomics is an emerging field that extracts large numbers of quantitative features from medical images, capturing aspects of tumor texture, shape, and heterogeneity that are invisible to the human eye. These features can then be used to train predictive models for clinical endpoints such as survival, treatment response, and recurrence.
Combined with artificial intelligence (AI) approaches including machine learning and deep learning, radiomics has the potential to transform how PDAC is evaluated by imaging, enabling more objective, reproducible, and personalized assessments than conventional radiology alone.
A complete radiomics workflow involves several sequential steps. It begins with image preparation, including quality checking, voxel resampling to a uniform resolution, and intensity normalization to ensure that features extracted from different scanners and protocols are comparable.
Tumor segmentation is one of the most critical and time-consuming steps. Manual segmentation by radiologists is currently the most common approach but is subject to inter-observer variability. Semi-automated and automated deep learning segmentation methods are being developed to improve reproducibility and scalability.
Feature extraction encompasses multiple categories: shape features (size, volume, surface area), first-order statistical features (intensity mean, variance, kurtosis), and second-order texture features including Gray Level Co-occurrence Matrix (GLCM) and Gray Level Run Length Matrix (GLRLM) metrics. Hundreds to thousands of features can be extracted from a single image.
Dimensionality reduction is essential because the number of features typically far exceeds the number of patients in the training set, creating high risk of overfitting. Techniques such as LASSO, principal component analysis (PCA), and recursive feature elimination are used to retain only the most predictive features.
Model development and external validation complete the pipeline. Internal cross-validation is insufficient alone; external validation on independent patient cohorts from different institutions is necessary to assess true generalizability and clinical utility.
One of the most active AI applications in PDAC imaging is automated tumor segmentation. Deep learning models, particularly U-Net architectures, have shown promising performance in delineating PDAC on CE-CT, though challenging cases with poor contrast enhancement or peritumoral inflammation remain difficult.
Radiomics features have been used to predict overall survival (OS) in PDAC patients. Multiple studies have identified texture features reflecting tumor heterogeneity as significant OS predictors independent of standard clinical factors, suggesting that imaging can capture biological variation not apparent from stage alone.
A clinically important application is predicting response to neoadjuvant chemotherapy before surgery. If imaging can identify patients unlikely to respond to chemotherapy, they might be triaged to alternative regimens or immediate surgery, avoiding the toxicity of ineffective treatment.
Radiomics has also been explored for distinguishing PDAC from benign pancreatic conditions such as autoimmune pancreatitis, which can mimic PDAC on imaging. Accurate non-invasive discrimination could spare patients unnecessary surgery or biopsy.
A major barrier to clinical translation of radiomics is the lack of standardization in how features are computed. Even when using the same image and the same software package, small differences in implementation can produce substantially different feature values, making results from different studies impossible to compare directly.
The Image Biomarker Standardisation Initiative (IBSI) was established to address this problem by defining reference standards for feature computation. Studies that comply with IBSI guidelines produce features that can in principle be compared and pooled across different software platforms and institutions.
However, many published radiomics studies were conducted before IBSI guidelines were established or do not report compliance, limiting their comparability. This is a recognized limitation in the field and is being addressed through updated guidelines, open-source software libraries, and mandatory reporting checklists such as TRIPOD.
Image harmonization methods, including ComBat and deep learning-based normalization, are being developed to retrospectively reduce scanner-related and protocol-related batch effects in multi-institutional radiomics datasets, enabling pooling of data from heterogeneous sources.
Most published PDAC radiomics studies have relatively small sample sizes, often fewer than 200 patients. Small datasets increase the risk of overfitting, where the model memorizes training data patterns rather than learning generalizable associations, leading to inflated performance estimates.
Tumor segmentation variability remains a source of feature instability. Different radiologists contouring the same tumor can produce meaningfully different segmentation volumes, and features sensitive to segmentation boundaries will be correspondingly unstable. Stable features that are robust to minor segmentation differences must be prioritized.
The black-box nature of many deep learning models makes it difficult to explain predictions to clinicians or regulators, hindering adoption. Explainability methods and radiological anchoring of predictions to interpretable image features are active research areas.
Prospective clinical trials that randomize patients to radiomics-guided versus standard management are needed to demonstrate that radiomics improves patient outcomes, not just statistical prediction metrics. None have yet been completed in PDAC.
The future of radiomics in PDAC will likely involve multicenter collaborative efforts that pool data from many institutions to build larger, more generalizable datasets. Initiatives modeled on TCGA and TCIA for imaging data could accelerate this process substantially.
Multimodal integration combining radiomics with genomic, proteomic, and clinical data holds particular promise. Imaging-genomics approaches may reveal how tumor genotype influences imaging phenotype, potentially creating richer predictive models than imaging or genomics alone.
The regulatory pathway for AI-based imaging biomarkers is becoming clearer with guidance from the FDA and CE mark frameworks. Prospective validation studies with clinical outcome endpoints will be required for regulatory approval and eventual reimbursement.
If these challenges are overcome, radiomics and AI have the potential to transform the management of PDAC by enabling non-invasive precision oncology: tailoring treatment decisions to the individual tumor's biology as revealed by the imaging it produces.