[Paper Review] Detecting Pancreatic Ductal Adenocarcinoma in Multi-phase CT Scans via Alignment Ensemble

Yingda Xia, Qihang Yu+4 more|arXiv (Cornell University)|Mar 18, 2020

Radiomics and Machine Learning in Medical Imaging24 references4 citations

TL;DR

This paper proposes an alignment ensemble method for detecting pancreatic ductal adenocarcinoma (PDAC) in multi-phase CT scans by combining early (image registration), late (feature space transformation), and slow (multi-level feature alignment) strategies. The ensemble significantly improves tumor segmentation DSC by 4% over state-of-the-art methods, achieving SOTA performance on two PDAC datasets through complementary feature-space alignment in an end-to-end deep learning framework.

ABSTRACT

Pancreatic ductal adenocarcinoma (PDAC) is one of the most lethal cancers among the population. Screening for PDACs in dynamic contrast-enhanced CT is beneficial for early diagnosis. In this paper, we investigate the problem of automated detecting PDACs in multi-phase (arterial and venous) CT scans. Multiple phases provide more information than single phase, but they are unaligned and inhomogeneous in texture, making it difficult to combine cross-phase information seamlessly. We study multiple phase alignment strategies, i.e., early alignment (image registration), late alignment (high-level feature registration), and slow alignment (multi-level feature registration), and suggest an ensemble of all these alignments as a promising way to boost the performance of PDAC detection. We provide an extensive empirical evaluation on two PDAC datasets and show that the proposed alignment ensemble significantly outperforms previous state-of-the-art approaches, illustrating the strong potential for clinical use.

Motivation & Objective

• To improve automated detection of pancreatic ductal adenocarcinoma (PDAC) in multi-phase CT scans, which are often underutilized due to misalignment and texture variation across phases.
• To address the challenge of combining information from arterial and venous phase CT scans, which are inherently unaligned due to patient motion and subtle tumor texture changes.
• To investigate and compare multiple phase alignment strategies—early, late, and slow alignment—for enhancing PDAC detection performance.
• To demonstrate that an ensemble of these alignment strategies yields superior performance by leveraging their complementary strengths.

Proposed method

• Employs a dual-encoder architecture, with one encoder per phase (arterial and venous), to extract features from multi-phase CT scans.
• Applies early alignment via image registration in the input space to align anatomical structures before feature extraction.
• Implements late alignment by performing spatial transformations on high-level features in the final feature space to align representations across phases.
• Introduces slow alignment as a multi-level feature registration strategy that progressively aligns features at different network depths.
• Combines all three alignment strategies into an ensemble model that aggregates predictions from individual alignment branches.
• Uses a segmentation decoder to generate PDAC tumor masks, optimizing for Dice Similarity Coefficient (DSC) and detection sensitivity.

Experimental results

Research questions

• RQ1 Can early image registration effectively improve PDAC detection in multi-phase CT scans despite motion-induced misalignment?
• RQ2 Do late and slow feature-space alignment strategies outperform image-space registration in terms of segmentation accuracy and robustness?
• RQ3 Are early, late, and slow alignment strategies complementary, such that their ensemble yields superior performance?
• RQ4 Can end-to-end feature-space alignment jointly optimize alignment and segmentation better than pre-processing registration?
• RQ5 Does the ensemble of alignment strategies reduce false negatives and improve sensitivity for early-stage PDAC detection?

Key findings

• The alignment ensemble model achieves a tumor DSC score of 64.14% on dataset ×2, representing a 4% improvement over the best individual alignment method (62.02% for late alignment).
• Late and slow feature-space alignment strategies outperform early image registration in segmentation performance, with DSC scores of 62.02% and 61.24% respectively, compared to 60.87% for early alignment.
• The ensemble model reduces the number of missed tumor detections from 8 to 6 on dataset ×1, improving sensitivity from 94.1% to 97.1%.
• Feature-space alignment methods (late and slow) yield larger performance gains than early alignment, as they jointly optimize alignment and segmentation via the final loss function.
• The three alignment strategies are complementary: combining them via ensemble learning leads to significant performance gains, demonstrating that no single alignment method dominates across all aspects of detection.
• The proposed method achieves state-of-the-art performance on both PDAC datasets, confirming the effectiveness of the alignment ensemble approach for multi-phase CT-based PDAC detection.