From Global Radiomics to Parametric Maps: A Unified Workflow Fusing Radiomics and Deep Learning for PDAC Detection

This paper presents a unified framework that fuses handcrafted radiomic features with deep learning by injecting them at both global and voxel levels into an nnUNet architecture, achieving state-of-the-art performance in pancreatic ductal adenocarcinoma detection on the PANORAMA dataset and an external cohort.

The Gist

Imagine you are trying to find a specific, tricky needle in a massive haystack. In the medical world, that "needle" is Pancreatic Ductal Adenocarcinoma (PDAC), a type of pancreatic cancer, and the "haystack" is a 3D CT scan of a patient's body.

For a long time, doctors and computers have used two different strategies to find this needle:

1. The "Big Picture" Detective (Radiomics): This approach looks at the whole haystack and measures its texture, shape, and density. It's like saying, "This pile of hay feels rougher and more uneven than a normal pile, so there's probably a needle in there." It's great at spotting the presence of a problem but bad at telling you exactly where it is.
2. The "Pixel-by-Pixel" Detective (Deep Learning): This is a super-smart AI that looks at every single grain of hay (voxel) in the image. It's amazing at drawing a precise outline around the needle, but sometimes it gets confused by the noise of the hay or gets tricked by how different scanners take pictures.

The Problem: Most previous attempts to combine these two detectives just took the "Big Picture" notes and handed them to the "Pixel" detective at the very end. It was like giving a map to a driver only after they had already finished driving the whole route. It didn't help them navigate the tricky turns along the way.

The New Solution: A Unified Workflow

The authors of this paper, a team from Cedars-Sinai and UCLA, built a unified framework that makes these two detectives work together from the very start. Think of it as a high-tech, two-stage search mission:

Stage 1: The Rough Sketch (Finding the Haystack)

First, the system uses a standard AI to quickly scan the whole body and find the pancreas (the haystack). It doesn't need to be perfect; it just needs to know, "Okay, the pancreas is roughly in this box."

Stage 2: The Precision Hunt (Finding the Needle)

Once the box is found, the system zooms in for a close-up. This is where the magic happens. They don't just feed the image to the AI; they feed it three things at once:

1. The Image: The actual CT scan.
2. The "Heat Maps" (Parametric Maps): This is the creative part. Instead of just looking at the image, the system calculates specific "texture scores" for every single pixel and turns them into colorful maps.
   • Analogy: Imagine the CT scan is a black-and-white photo of a forest. The system creates a second photo where every tree that looks "suspiciously rough" glows red, and every tree that looks "smooth" glows blue. Now the AI isn't just looking at a photo; it's looking at a photo plus a glowing map of suspicious spots.
3. The "Cheat Sheet" (Global Features): The system also remembers the "Big Picture" clues it found earlier (like the overall shape of the organ) and whispers them to the AI right when it's making its hardest decisions.

Why This Works So Well

The paper tested this method on a huge dataset called PANORAMA. Here's what happened:

• The Baseline: A standard AI (nnUNet) was good, but it missed some tricky cases.
• The Upgrade: When they added the "Heat Maps" and the "Cheat Sheet," the AI became much sharper.
  • On the main test, it got a score of 0.96 (out of 1.0), which is nearly perfect.
  • On a completely new, unseen group of patients (the "external" test), it still scored 0.95, proving it wasn't just memorizing the answers but actually learning the rules.
• The Competition: This method was so good that it took 2nd place in a major international competition (the PANORAMA Grand Challenge), beating out many other teams.

The Secret Sauce: Speed and Clarity

Usually, creating these "Heat Maps" for every single pixel takes forever (like trying to count every grain of sand on a beach one by one). The authors wrote a special, super-fast computer program (using GPUs) that does this calculation in seconds instead of minutes. This makes the whole process practical for real hospitals.

The Takeaway

This paper shows that the best way to find a needle in a haystack isn't to choose between a "Big Picture" view or a "Close-up" view. Instead, you should combine them.

By giving the AI a "heat map" of suspicious textures and a summary of the overall shape, they created a system that is:

• More Accurate: It finds more cancers.
• More Robust: It doesn't get confused by different types of CT scanners.
• More Efficient: It runs fast enough to be used in real life.

In short, they taught the AI to look at the forest and the trees simultaneously, using a special set of glowing maps to guide its way.

Technical Summary

1. Problem Statement

Pancreatic Ductal Adenocarcinoma (PDAC) detection via CT imaging is critical but challenging. While Deep Learning (DL) models (specifically nnUNet) have achieved high performance, they often suffer from:

• Lack of Interpretability: Purely data-driven models act as "black boxes."
• Domain Shift: Vulnerability to variations across different scanners and medical sites.
• Underutilization of Radiomics: Existing fusion methods typically rely on global radiomic vectors (summary statistics of the whole tumor) and ignore spatially resolved radiomic parametric maps.
• Computational Barriers: Generating voxel-wise parametric maps for many descriptors is computationally expensive, often leading researchers to compress maps (e.g., via SVD), which reduces feature interpretability.

The authors aim to bridge the gap between handcrafted radiomics (interpretable, reproducible) and deep learning (powerful feature extraction) by creating a unified workflow that leverages both global biomarkers and voxel-level spatial cues.

2. Methodology

The proposed framework is a two-stage, radiomics-aware nnUNet pipeline designed for PDAC detection.

A. Data Preparation

• Datasets:
  • Training: PANORAMA Challenge dataset (2,238 venous-phase contrast-enhanced CT scans).
  • External Validation: Internal Cedars-Sinai Medical Center cohort (218 scans).
• Preprocessing: 5-fold stratified cross-validation based on lesion size (voxels) to ensure balanced distribution.

B. Step 1: Global Radiomics Analysis & Feature Selection

Before training the DL model, a global analysis is performed to identify the most discriminative features:

1. Extraction: 1,486 radiomic features (first-order, texture families like GLCM/GLRLM, and shape) are extracted from the whole pancreas mask using PyRadiomics.
2. Selection Pipeline:
   • Univariate Filtering: Retains features with significant Pearson correlation between PDAC and non-PDAC groups (FDR controlled).
   • Recursive Feature Elimination (RFE): An SVM is used to iteratively eliminate features until only the top 10 remain.
3. Final Set: The top 10 features are selected. Shape descriptors are excluded for the parametric map generation, leaving 8 features to be converted into voxel-wise parametric maps.

C. Step 2: Two-Stage nnUNet Architecture

The detection pipeline consists of two cascaded networks:

• Stage 1: Coarse Localization

  • Input: Low-resolution CT volumes.
  • Task: Predicts a rough pancreas mask to define a Region of Interest (ROI).
  • Output: A fixed 100 × 50 × 15 mm ROI cropped from the original high-resolution CT.
  • Note: Radiomics for Stage 2 are computed based on this predicted mask (not ground truth) to simulate real-world deployment.

• Stage 2: Fine Segmentation/Detection

  • Input: High-resolution cropped ROI.
  • Multi-Structure Supervision: The model is trained to jointly predict the pancreas, abdominal aorta, portal vein, pancreatic duct, and common bile duct (not just PDAC) to improve anatomical context and robustness.
  • Radiomics Fusion Mechanisms:
    1. Voxel-Level (Parametric Maps): The 8 selected radiomic features are converted into voxel-wise parametric maps. These are concatenated channel-wise with the CT input at the network's input layer.
    2. Global-Level (Case Vectors): The global radiomic vector (from the 10 selected features) is injected at the bottleneck of the nnUNet via Multi-Head Cross-Attention. Here, the global vector acts as the query, and the latent nnUNet features act as keys/values.
  • Loss Function: Combination of Dice loss and normalized cross-entropy.

D. Technical Innovation: CUDA-Accelerated Extraction

To make voxel-wise radiomics feasible for large cohorts, the authors implemented a GPU-accelerated radiomics extractor using PyTorchRadiomics. This replaces the CPU-based vanilla PyRadiomics, drastically reducing computation time.

3. Key Contributions

1. Unified Radiomics-DL Workflow: A novel approach that uses global radiomics not just as a final classifier, but as a selection signal to pinpoint a compact subset of features. These features are then materialized as parametric maps (spatial) and global embeddings (contextual) for the DL model.
2. Radiomics-Enhanced nnUNet: The architecture fuses radiomics at two distinct levels:
   • Input Level: Parametric maps provide spatial priors.
   • Bottleneck Level: Cross-attention integrates global case-level context.
3. Scalable Implementation: The use of GPU-accelerated parametric map extraction reduces feature extraction time from ~53s to ~16s per feature, enabling large-scale parametric map generation.

4. Experimental Results

The method was evaluated on the PANORAMA dataset (5-fold CV) and an external Cedars-Sinai cohort.

• Performance Metrics (AUC / Average Precision):
  • PANORAMA (CV):
    • Baseline nnUNet: AUC 0.959 / AP 0.810
    • + Both (Global + Maps): AUC 0.958 / AP 0.836 (Highest combined score).
  • External Cohort:
    • Baseline nnUNet: AUC 0.954 / AP 0.662
    • + Both: AUC 0.951 / AP 0.777 (Statistically significant improvement, $p < 0.01$).
• Ablation Study:
  • Adding Parametric Maps alone provided a larger boost than adding Global Features alone. The authors attribute this to the "noisy" global information resulting from imperfect Stage 1 segmentation, whereas parametric maps offer direct spatial guidance.
• Competition Ranking:
  • The model ranked 2nd in the PANORAMA Grand Challenge. Notably, the preliminary submission used only parametric maps (without global fusion) yet still achieved a top-tier ranking.
• Runtime:
  • GPU acceleration reduced mean extraction time per feature from 53.42s to 16.28s ($p \ll 0.01$).

5. Significance

• Complementary Signals: The study demonstrates that handcrafted radiomics, when injected at both global and voxel levels, provide complementary signals that enhance the sensitivity and robustness of deep learning models for PDAC detection.
• Interpretability: By selecting specific discriminative features and visualizing them as parametric maps, the model bridges the gap between "black box" DL and interpretable radiomics.
• Clinical Scalability: The GPU-accelerated extraction pipeline solves the computational bottleneck of voxel-wise radiomics, making this unified workflow practical for real-world, large-scale clinical applications.
• Generalizability: The method's success on an external, in-house cohort suggests it is robust against domain shifts (different scanners/sites), a common failure point for pure DL models.