The FELIX Project: Deep Networks To Detect Pancreatic Neoplasms

Yingda Xia, Qihang Yu, Linda C. Chu, Satomi Kawamoto, Seyoun Park, Fengze Liu, Jieneng Chen, Zhuotun Zhu, Bowen Li, Zongwei Zhou, Yongyi Lu, Yan Wang, Wei Shen, Lingxi Xie, Yuyin Zhou, Christopher L. Wolfgang, Ammar A. Javed, Daniel Fadaei Fouladi, Shahab Shayesteh, Jefferson S. Graves, Alejandra Blanco, Eva S. Zinreich, Miriam Klauß, Philipp Mayer, Benedict Kinny‐Köster, Kenneth W. Kinzler, Ralph H. Hruban, Bert Vogelstein, Alan Yuille, Elliot K. FishmanmedRxiv 2022 · DOI

Abstract

Tens of millions of abdominal images are obtained with computed tomography (CT) in the U.S. each year, but pancreatic cancers are sometimes not initially detected. We describe a suite of algorithms, named FELIX, that can recognize pancreatic lesions from CT images without human input. Using FELIX, >95% of patients with pancreatic ductal adenocarcinomas (PDAC) were detected at a specificity of >95% in patients without pancreatic disease. FELIX may assist radiologists in identifying pancreatic cancers earlier, when surgery and other treatments offer more hope for long-term survival.

Introduction: The Challenge of Early Detection

Pancreatic cancer remains one of the most lethal malignancies, largely because it is often diagnosed at an advanced, unresectable stage. While over 40 million abdominal CT scans are performed annually in the U.S., early-stage lesions are frequently subtle and can be easily overlooked by radiologists, particularly when scans are ordered for unrelated clinical concerns.

The FELIX Project addresses this critical gap by introducing a suite of deep learning algorithms designed to act as a "second reader." The technical hurdles are significant: the pancreas exhibits high anatomical variability in shape and position, and AI models often struggle with "domain transfer"—maintaining performance across different hospital systems and imaging hardware.

Clinical examples of missed vs. detected pancreatic cancer and the FELIX system's 2D/3D visualization capabilities.

Methods: The FELIX Architecture

The researchers developed FELIX using a modified 3D U-Net architecture, evolving through several iterations to handle clinical complexities:

  • Auto-Alignment and Dual-Phase Imaging: To maximize diagnostic yield, FELIX utilizes both arterial and venous phase CT scans. A novel Auto-Alignment Module, discovered via Neural Architecture Search (NAS), automatically aligns and fuses these images to account for physical offsets between phases.
  • Two-Stage Detection (FELIX 1.1): The system first localizes the pancreas within a bounding box before performing voxel-wise segmentation to distinguish healthy tissue from tumors.
  • Robustness through Augmentation (FELIX 1.2): To overcome the domain transfer problem, the model was trained with aggressive data augmentation, simulating various 3D rotations and slice thicknesses to mimic real-world clinical diversity.
  • Multiscale Processing (FELIX 1.3): This iteration employs multiscale techniques to improve the detection of small lesions, such as cysts and pancreatic neuroendocrine tumors (PanNETs).
  • False Positive Reduction: A post-processing stage filters noise by evaluating features like sphericity, entropy-based uncertainty, and Variational Autoencoder (VAE) reconstruction error.
The FELIX project workflow, illustrating the training phase with manual annotations and the testing phase for quantitative evaluation and radiologist review.

Datasets and Validation

The study utilized a massive, curated dataset of over 2,500 dual-phase contrast-enhanced CT scans. The primary training and testing were conducted on a cohort from Johns Hopkins Hospital (JHH), including 836 healthy controls and over 1,300 cases of PDAC, PanNETs, and cysts.

Crucially, the model was validated against external datasets to ensure generalizability:

  • Multi-Center U.S. Validation: 399 PDAC patients from various hospitals using GE, Siemens, Philips, and Toshiba scanners.
  • International Validation: 242 dual-phase scans from Heidelberg Medical School.

Key Results and Insights

FELIX demonstrated expert-level performance across multiple metrics:

  • High Accuracy: On the JHH test set, the system achieved 97% sensitivity at 99% specificity for PDAC detection.
  • Generalization: Sensitivity remained above 97% on external multi-center data, proving the model's robustness across different imaging protocols.
  • Small Lesion Detection: The model successfully detected PDACs and cysts smaller than 2cm with approximately 77% and 76% sensitivity, respectively.
  • AI as a Corrective Tool: During re-review, FELIX identified several tumors that had been missed by human annotators. Cases initially flagged as "false positives" were often revealed to be true positives upon expert re-evaluation.

Conclusion and Future Directions

The FELIX Project demonstrates that deep learning can provide a highly reliable automated screening tool for pancreatic neoplasms. While the system shows remarkable sensitivity, challenges remain in maintaining high localization accuracy across diverse external datasets. Additionally, future work will focus on differentiating neoplasms from inflammatory conditions like pancreatitis. The next phase involves large-scale prospective clinical trials to integrate FELIX into real-time diagnostic workflows, potentially transforming the early detection landscape for pancreatic cancer.

Summary updated

Abstract

ABSTRACT Tens of millions of abdominal images are obtained with computed tomography (CT) in the U.S. each year but pancreatic cancers are sometimes not initially detected in these images. We here describe a suite of algorithms (named FELIX) that can recognize pancreatic lesions from CT images without human input. Using FELIX, > 95% of patients with pancreatic ductal adenocarcinomas were detected at a specificity of > 95% in patients without pancreatic disease. FELIX may be able to assist radiologists in identifying pancreatic cancers earlier, when surgery and other treatments offer more hope for long-term survival.
BibTeX
@misc{xia2022felix,
  author = {Yingda Xia and Qihang Yu and Linda C. Chu and Satomi Kawamoto and Seyoun Park and Fengze Liu and Jieneng Chen and Zhuotun Zhu and Bowen Li and Zongwei Zhou and Yongyi Lu and Yan Wang and Wei Shen and Lingxi Xie and Yuyin Zhou and Christopher L. Wolfgang and Ammar A. Javed and Daniel Fadaei Fouladi and Shahab Shayesteh and Jefferson S. Graves and Alejandra Blanco and Eva S. Zinreich and Miriam Klauß and Philipp Mayer and Benedict Kinny‐Köster and Kenneth W. Kinzler and Ralph H. Hruban and Bert Vogelstein and Alan Yuille and Elliot K. Fishman},
  title = {{The FELIX Project: Deep Networks To Detect Pancreatic Neoplasms}},
  year = {2022},
  doi = {10.1101/2022.09.24.22280071},
  howpublished = {medRxiv},
}
Figure