CMFCUNet: cascaded multi-scale feature calibration UNet for pancreas segmentation

Chengjian Qiu, Yuqing Song, Zhe Liu, Jing Yin, Kai Han, Yi Liu

Abstract

Segmenting the pancreas from abdominal CT scans is challenging since it often takes up a relatively small region. Researchers have suggested leveraging coarse-to-fine approaches to cope with this challenge. However, the coarse-scaled segmentation and the fine-scaled segmentation are typically either trained separately, utilizing coordinates located by the coarse mask to crop the input, or trained jointly, utilizing the mask to enhance the input. This paper argues that these two solutions are complementary and proposes a cascaded multi-scale feature calibration UNet (CMFCUNet). In this framework, multi-scale features are calibrated vertically in a pixel-wise fashion, and the coarse and fine stages are connected via a Dual Enhancement Module (DEM). Experiments on the public NIH pancreas dataset show that CMFCUNet increases the Jaccard index (JI) by over 3% and the Dice Similarity Coefficient (DSC) by nearly 1%, surpassing existing approaches while remaining highly parameter-efficient.

What Problem Does This Paper Address?

Automated pancreas segmentation is a classic "needle in a haystack" problem in medical imaging. The pancreas is a small organ, often accounting for less than 0.5% of the total volume of an abdominal CT scan. Furthermore, it possesses highly irregular boundaries that frequently blend into surrounding tissues, making precise localization and segmentation difficult for standard deep learning models.

Visualization of pancreas segmentation (red) in NIH Case#010 across axial, sagittal, and coronal views, highlighting the organ's small size and complex boundaries.

To address this, researchers typically employ "coarse-to-fine" strategies. A coarse model first identifies the general region of interest, and a second "fine" model performs detailed segmentation. However, existing methods usually follow one of two distinct paths:

  1. Separate Training with Cropping: The coarse model provides coordinates to crop the image for the fine model. While this removes background noise, it can result in the loss of global context.
  2. Joint Training with Enhancement: The coarse model's output acts as a weight or additional channel to highlight the pancreas. This maintains context but often fails to fully suppress irrelevant background regions.
Comparison of traditional coarse-to-fine strategies: (a) cropping-based methods and (b) joint-training enhancement methods.

Additionally, standard UNet architectures often suffer from feature redundancy. They pass raw data from the encoder to the decoder via skip connections without filtering or calibrating the data, which can introduce noise into the final segmentation.

Methods

The authors propose CMFCUNet, which introduces two primary architectural innovations designed to overcome the limitations of current coarse-to-fine frameworks.

Dual Enhancement Module (DEM)

Instead of choosing between cropping and weighting, the DEM integrates both. It adaptively transforms the coarse segmentation map to multiply (enhance) the input image and then crops the result based on the coarse model's localization. This ensures the fine-scaled network receives a "cleaned" and "zoomed-in" version of the data, benefiting from both local focus and contextual guidance.

Multi-scale Feature Calibration Gate (MFCG)

Within the UNet backbone (referred to as MFCUNet), the authors replaced standard skip connections with a calibration gate. Using a mechanism similar to attention (Query, Key, Value), the gate allows higher-level semantic features to "query" lower-level high-resolution features. This vertically calibrates the features pixel-by-pixel, effectively filtering out noise and redundancy before the data reaches the decoder.

Datasets and Evaluation

The study utilized the NIH pancreas segmentation dataset, a standard public benchmark consisting of 82 contrast-enhanced abdominal CT volumes. The authors followed a rigorous 4-fold cross-validation protocol to ensure the reliability of their results. Performance was measured using the Dice Similarity Coefficient (DSC) and the Jaccard Index (JI).

Key Results

CMFCUNet set a new performance benchmark on the NIH dataset:

  • Accuracy: It achieved a DSC of 86.30% and a JI of 76.26%, outperforming existing mainstream coarse-to-fine approaches.
  • Ablation Improvements: The vertical calibration (MFCG) alone improved the Jaccard Index by over 4% compared to baseline models.
  • Efficiency: The model is remarkably lightweight, using only 13.68 million parameters. In comparison, some state-of-the-art models use over 250 million parameters while achieving lower accuracy.
  • Speed: The average inference time was recorded at 1.26 minutes per case, supporting its potential for clinical workflows.

Conclusion

CMFCUNet demonstrates that the synergy between cropping and feature enhancement, combined with vertical feature calibration, significantly advances the state-of-the-art in pancreas segmentation. By focusing on architectural intelligence rather than brute-force scaling, the model achieves superior accuracy with high parameter efficiency. Future work may involve extending the framework to 3D networks and testing on larger, more diverse clinical datasets to ensure broader generalizability.

Summary updated

BibTeX
@inproceedings{qiu2022cmfcunet,
  author = {Chengjian Qiu and Yuqing Song and Zhe Liu and Jing Yin and Kai Han and Yi Liu},
  title = {{CMFCUNet: cascaded multi-scale feature calibration UNet for pancreas segmentation}},
  booktitle = {Research},
  year = {2022},
  doi = {10.21203/rs.3.rs-1954951/v1},
}
Figure