PGD-UNet: A Position-Guided Deformable Network for Simultaneous Segmentation of Organs and Tumors

Ziqiang Li, Hong Pan, Yaping Zhu, A. Kai Qin

TL;DR

This paper introduces PGD-UNet, a novel deep learning architecture for segmenting organs and tumors in medical images. It addresses challenges like irregular shapes, size variations, and class imbalance by incorporating position-guided deformable convolutions and a new noise-suppression loss function, achieving state-of-the-art results on bladder and pancreas segmentation tasks.

What Problem Does This Paper Address?

Accurate segmentation of organs and tumors in medical images is critical for diagnosis and treatment planning. However, this task is fraught with difficulties:

  • Irregular Shapes and Size Variations: Tumors, in particular, can have highly irregular shapes and vary significantly in size depending on the cancer stage. Organs also exhibit natural variations.
  • Class Imbalance: The region of interest (e.g., a tumor) often occupies a tiny fraction of the image compared to the background, making it hard for models to learn effectively.
  • Overlapping Structures: Tumors frequently overlap with surrounding organs, making it challenging to delineate them precisely and simultaneously.
  • Label Noise: Medical image annotations can be subjective, and image quality can vary, leading to inaccuracies in the ground truth labels. This "label noise" can cause models to overfit and perform poorly.

Existing methods often struggle to handle all these challenges simultaneously, especially when segmenting both organs and tumors in a single pass.

Methods

The core of the proposed solution is the PGD-UNet, which builds upon the successful U-Net architecture. Here's a breakdown of its key innovations:

  • Position-Guided Deformable Convolution (PGD-Conv): Standard deformable convolutions learn offsets to adapt their sampling grid to object shapes. PGD-UNet enhances this by explicitly feeding position information into the deformable convolution operator. This is achieved through an auxiliary localization path that generates coordinate information.
  • CoordConv: This technique injects pixel coordinates into the feature maps, making the network aware of spatial locations.
  • CoordPool: A novel pooling module that preserves the spatial location of the maximum value within a pooling region, addressing the loss of positional information in traditional max-pooling.
  • Atrous Spatial Pyramid Pooling (ASPP): Used in the bottleneck of the U-Net, ASPP employs dilated convolutions with different rates to capture multi-scale features. This helps the network represent structures of various sizes effectively.
  • Noise Suppression Focal Loss (NSFL): To combat label noise, the authors propose NSFL. It's a modification of Focal Loss, which already down-weights easy examples. NSFL introduces a piecewise modulating factor that further suppresses the contribution of potentially mislabeled samples (those with low predicted probabilities) to the loss, preventing overfitting.
  • Combined Loss Function: The final loss is a weighted combination of NSFL and Dice Loss, balancing the benefits of both for segmentation accuracy and handling class imbalance.

Datasets

The PGD-UNet was evaluated on two challenging medical segmentation tasks:

  • Bladder Tumor Dataset: Consists of 2200 high-resolution Axial T2-weighted MRI slices from 25 patients with bladder cancer. Both bladder wall and tumor regions were manually delineated.
  • Pancreas Tumor Dataset: A sub-dataset from the Medical Segmentation Decathlon (MSD) challenge, comprising 282 portal venous phase CT scans. Expert radiologists annotated pancreatic parenchyma and pancreatic mass.

Key Insights and Ideas

  • Position Matters for Deformable Convolutions: The paper strongly suggests that explicitly guiding deformable convolutions with positional information (via CoordConv and CoordPool) significantly improves their ability to model complex geometric transformations, especially for challenging structures like tumors.
  • Max-Pooling Loses Crucial Spatial Cues: The introduction of CoordPool highlights how standard max-pooling discards valuable location information that can be detrimental for precise segmentation.
  • Label Noise is a Major Hurdle: The NSFL demonstrates that directly applying standard loss functions can be problematic when dealing with noisy labels, leading to poor generalization. A tailored loss function is crucial for robust training.
  • Tumors are the Real Challenge: Ablation studies show that the proposed PGD-UNet modules (especially the localization path and NSFL) provide a more substantial improvement for tumor segmentation compared to normal tissue segmentation, underscoring their effectiveness in handling irregular shapes and potential label noise associated with tumors.

Key Results

  • State-of-the-Art Performance: PGD-UNet achieved state-of-the-art results on both the bladder and pancreas tumor segmentation tasks.
  • Significant Tumor Segmentation Improvement: For bladder tumors, PGD-UNet achieved a Dice score of 80.38%, a substantial improvement over baseline methods. For pancreas tumors, it reached 50.12%.
  • Competitive with 3D Methods: The 2D PGD-UNet achieved comparable accuracy to more complex 3D methods while requiring less computational power and simpler annotation.
  • Improved Accuracy over Baselines: Compared to a standard U-Net, PGD-UNet showed moderate improvement in bladder wall segmentation and significant gains in bladder tumor segmentation.
  • Effective Noise Suppression: The NSFL component consistently improved tumor segmentation DSC, while slightly degrading normal tissue performance, indicating its targeted benefit for noisy tumor labels.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while NSFL helps with label noise, it might slightly degrade the performance on normal tissues.
  • The current work focuses on 2D segmentation, and while competitive, 3D methods might still offer advantages in certain scenarios.
  • Future Work:
    • The authors plan to extend their work to incorporate complementary information from both MRI and CT images, which will require addressing registration challenges.

Summary updated

Abstract

Precise segmentation of organs and tumors plays a crucial role in clinical applications. It is a challenging task due to the irregular shapes and various sizes of organs and tumors as well as the significant class imbalance between the anatomy of interest (AOI) and the background region. In addition, in most situation tumors and normal organs often overlap in medical images, but current approaches fail to delineate both tumors and organs accurately. To tackle such challenges, we propose a position-guided deformable UNet, namely PGD-UNet, which exploits the spatial deformation capabilities of deformable convolution to deal with the geometric transformation of both organs and tumors. Position information is explicitly encoded into the network to enhance the capabilities of deformation. Meanwhile, we introduce a new pooling module to preserve position information lost in conventional max-pooling operation. Besides, due to unclear boundaries between different structures as well as the subjectivity of annotations, labels are not necessarily accurate for medical image segmentation tasks. It may cause the overfitting of the trained network due to label noise. To address this issue, we formulate a novel loss function to suppress the influence of potential label noise on the training process. Our method was evaluated on two challenging segmentation tasks and achieved very promising segmentation accuracy in both tasks.
BibTeX
@inproceedings{li2020pgd,
  author = {Ziqiang Li and Hong Pan and Yaping Zhu and A. Kai Qin},
  title = {{PGD-UNet: A Position-Guided Deformable Network for Simultaneous Segmentation of Organs and Tumors}},
  booktitle = {IJCNN},
  year = {2020},
  url = {https://openreview.net/forum?id=H3VujAfRg1},
}
Figure