Region-Manipulated Fusion Networks for Pancreatitis Recognition

Jian Wang, Xiaoyao Li, Xiangbo Shu, Weiqin LiarXiv 2019 · arXiv

Abstract

Acute Pancreatitis (AP) is a critical gastrointestinal condition where early and accurate diagnosis via Computed Tomography (CT) is vital. However, manual interpretation is subjective and labor-intensive, while standard deep learning models often fail to capture the fine-grained, non-rigid lesions characteristic of the disease. This paper introduces Region-Manipulated Fusion Networks (RMFN), a novel deep learning architecture specifically designed for AP recognition. RMFN employs a multi-network architecture that fuses global context from a main-network with local details from multi-scale sub-networks. Evaluated on a newly curated dataset of 5,000 CT images (ACIP), the proposed method achieves a peak accuracy of 96.7%, significantly outperforming traditional baselines and standard CNN architectures.

Introduction

Acute Pancreatitis is a common and potentially fatal gastrointestinal disorder. While Computed Tomography (CT) is the gold standard for diagnosis, the process is fraught with challenges. Radiologists' interpretations can vary based on experience, and manually reading CT slices is labor-intensive. Furthermore, early-stage lesions are often tiny ("fine-grained"), have no fixed shape ("non-rigid"), and look very similar to healthy tissue, leading to significant inter-class ambiguity.

Standard deep learning models typically process images globally. In medical imaging, however, the diagnostic signal (the lesion) might only occupy a tiny fraction of the image. This causes standard models to struggle as the critical features are often "washed out" by the surrounding healthy tissue and background noise.

The ACIP Dataset

To address the lack of public data for this task, the authors collected and released a new benchmark: the Abdominal CT Images about Pancreatitis (ACIP) dataset. It contains 5,000 real-world CT images collected from hospitals, balanced between 2,500 healthy (normal) images and 2,500 images from patients with pancreatitis. These images were annotated by three experienced doctors to ensure ground-truth accuracy.

Comparison of normal abdominal CT scans (a) and those showing signs of acute pancreatitis (b-d) with lesions highlighted in red.

Methodology: Region-Manipulated Fusion Networks

The core innovation of this work is the Region-Manipulated Fusion Network (RMFN). The architecture is designed to force the network to pay attention to local textures while maintaining global context through a multi-network scheme.

Multi-Network Architecture

The system consists of one Main-network capturing the global abdominal structure and two Sub-networks focusing on local details at different scales. The input image is divided into grids (e.g., $7 \times 7$ and $5 \times 5$), and these small regions are fed into the sub-networks.

The RMFN architecture, featuring a main-network fused with two multi-scale sub-networks for local feature extraction and ceaseless aggregation.

Region Pooling and Fusion

The authors utilize two region pooling strategies. "Region Pooling 1" creates a standard grid, while "Region Pooling 2" introduces overlapping areas. This overlap ensures that no lesion features are lost if they happen to fall on the boundaries of grid cells—a common failure point in patch-based recognition. The local feature maps from these sub-networks are then "fused" back into the main network's feature maps. This "ceaseless aggregation" reinforces lesion-related signals while weakening irrelevant background data.

Key Insights and Results

The RMFN approach relies on the insight that for medical diagnosis, local texture is often more discriminative than global shape. By using a multi-scale approach with different grid sizes, the network can capture non-rigid lesions of varying dimensions.

In performance evaluations, RMFN versions consistently outperformed their "vanilla" counterparts:

  • Top Performance: The AlexNet-RMFN-C model achieved the highest accuracy at 96.7%.
  • Architecture Boost: Integrating the RMFN scheme into VGG16 improved its accuracy from 90.6% to 94.7%.
  • Deep Learning Superiority: A traditional SVM (Support Vector Machine) approach only managed 78.6% accuracy, demonstrating the necessity of deep, multi-scale feature extraction for this complex task.

Visualization of feature maps confirmed that RMFN models localized actual lesion areas, such as pancreatic swelling or fluid collection, much more accurately than standard CNNs.

Conclusion and Future Work

While RMFN provides a significant leap in automated pancreatitis recognition, it does introduce higher computational overhead due to the parallel network structure. Future work will focus on improving generalization across different CT scanner brands and exploring Generative Adversarial Networks (GANs) to generate synthetic training data or further refine feature localization. This research represents a vital step toward reducing human subjectivity and improving efficiency in clinical pancreatitis diagnosis.

Summary updated

Abstract

This work first attempts to automatically recognize pancreatitis on CT scan images. However, different form the traditional object recognition, such pancreatitis recognition is challenging due to the fine-grained and non-rigid appearance variability of the local diseased regions. To this end, we propose a customized Region-Manipulated Fusion Networks (RMFN) to capture the key characteristics of local lesion for pancreatitis recognition. Specifically, to effectively highlight the imperceptible lesion regions, a novel region-manipulated scheme in RMFN is proposed to force the lesion regions while weaken the non-lesion regions by ceaselessly aggregating the multi-scale local information onto feature maps. The proposed scheme can be flexibly equipped into the existing neural networks, such as AlexNet and VGG. To evaluate the performance of the propose method, a real CT image database about pancreatitis is collected from hospitals \footnote{The database is available later}. And experimental results on such database well demonstrate the effectiveness of the proposed method for pancreatitis recognition.
BibTeX
@misc{wang2019region,
  author = {Jian Wang and Xiaoyao Li and Xiangbo Shu and Weiqin Li},
  title = {{Region-Manipulated Fusion Networks for Pancreatitis Recognition}},
  year = {2019},
  eprint = {1907.01744},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1907.01744},
}
Figure