Pancreatic ductal adenocarcinoma (PDAC) is characterized by poor survival even after surgical resection. Identifying which resected patients are at high versus low risk for early recurrence is critical for tailoring adjuvant chemotherapy and follow-up schedules.
Traditional prognostic tools rely on pathological staging and clinical variables, but these often fail to capture the full biological heterogeneity of PDAC. Both radiomics - quantitative feature extraction from CT images - and deep learning have independently shown promise for prognostic modeling, yet each approach captures different aspects of tumor biology.
This study from the University of Toronto investigated whether fusing radiomic and deep learning features using a novel risk-score-based method could outperform traditional feature combination strategies for predicting overall survival in resectable PDAC patients.
From preoperative CT scans, 1,428 radiomic features were extracted using the PyRadiomics library, covering first-order statistics, shape descriptors, and texture matrices including GLCM, GLRLM, and wavelet-derived features applied across multiple frequency bands.
In parallel, 35 deep learning features were extracted from the same CT images using LungTrans, a convolutional neural network pretrained on a large lung cancer dataset and applied via transfer learning. These features encode high-level semantic image representations that complement the hand-crafted radiomic descriptors.
Combining two feature sets of very different sizes and distributions required a principled feature fusion strategy. The researchers compared four traditional integration methods - PCA dimensionality reduction, Boruta feature selection, Cox Proportional Hazards (CPH) modeling, and LASSO regularization - against a novel risk-score-based fusion approach.
The proposed risk score-based fusion method works by first training separate survival models on the radiomics feature set and the deep learning feature set independently, generating a continuous risk score for each patient from each modality.
These two unimodal risk scores are then combined as a two-dimensional input into a final Random Forest ensemble model. This approach preserves the prognostic signal unique to each modality while avoiding the curse of dimensionality that arises from directly concatenating over 1,400 features.
All models were evaluated using the Area Under the Receiver Operating Characteristic curve (AUC) for binary survival classification (above/below median survival), with cross-validation used to estimate generalization performance and prevent overfitting on the relatively small PDAC cohort.
The risk score-based fusion method achieved an AUC of 0.84 for survival prediction, substantially higher than the best traditional feature integration approach, which reached an AUC of only 0.60 - representing a relative improvement of approximately 40%.
Among the traditional methods, LASSO-based feature selection performed best but still fell well short of the proposed fusion strategy. PCA and CPH-based methods performed similarly to each other but worse than LASSO, highlighting that naive dimensionality reduction fails to exploit complementary information across modalities.
The dramatic performance gap underscores that how radiomic and deep learning features are integrated matters as much as which features are extracted - risk-score compression preserves survival-relevant information while discarding redundant noise that confounds direct concatenation methods.
An AUC of 0.84 in a notoriously difficult-to-predict cancer like PDAC is clinically meaningful. A robust prognostic model could guide oncologists in selecting patients for intensified adjuvant chemotherapy regimens versus observation-only strategies after resection.
The use of preoperative CT images as the sole input means no additional biopsy, biomarker assay, or specialized imaging is required - the prognostic score is generated from imaging that is already part of standard surgical workup, enabling seamless clinical integration.
The risk-score-based fusion framework is modality-agnostic and could in principle be extended to incorporate additional data types such as MRI, PET, or genomic data, making it a generalizable architecture for multimodal oncology prognostication beyond PDAC.
The study cohort was relatively small, which is common in PDAC research given low resection rates. Small sample sizes risk overfitting even with cross-validation, and external validation on datasets from different institutions is essential before clinical deployment.
The LungTrans deep learning model was pretrained on lung cancer data rather than pancreatic cancer, which may limit the biological relevance of the extracted features. Pancreas-specific pretraining on a larger cohort could potentially improve the quality of deep features and further boost performance.
Future research should explore the biological interpretability of the fused risk score - understanding which imaging phenotypes drive high-risk classifications would build clinician trust and could reveal new targets for therapeutic intervention in resectable PDAC.