Automatic Lobe Segmentation Using Attentive Cross Entropy and End-to-End Fissure Generation

Xiaofan Zhang, Qi Su, Na Wang, Jiawen Xie, Yinan Chen, Xiaofan Zhang 0002ISBI 2023 · DOI

TL;DR

This paper introduces a novel deep learning framework for automatically segmenting lung lobes in CT scans. It tackles the challenge of incomplete fissures by using an "Attentive Cross Entropy" loss to focus on fissure regions and an end-to-end fissure generation module, achieving state-of-the-art results on public and private datasets.

What Problem Does This Paper Address?

Accurately segmenting individual lung lobes is crucial for diagnosing and treating lung diseases. However, this task is challenging in CT images due to the often incomplete or invisible nature of the pulmonary fissures that define lobe boundaries. These fissures are thin structures with subtle intensity differences from surrounding lung tissue, and their appearance can vary significantly across patients. Traditional methods struggle with these inconsistencies, leading to inaccurate lobe segmentation.

Methods

The proposed framework integrates several novel components:

  • Attentive Cross Entropy (ACE) Loss: This task-specific loss function is designed to make the model pay more attention to the challenging regions around pulmonary fissures. It achieves this by dynamically assigning higher weights to "hard" voxels (those that are misclassified or near boundaries), inspired by anchor loss. The attention weight is a dynamic value that increases during training to avoid destabilizing early learning.
  • End-to-End Fissure Generation Module (FGM): Instead of relying on separate, potentially less effective fissure segmentation branches, the FGM is integrated directly into the lobe segmentation network. It generates a fissure probability map by performing operations (max pooling and element-wise multiplication) on the predicted lobe probability maps, mimicking morphological dilation and intersection. This module is designed to be gradient-traceable, allowing the fissure supervision signal to propagate back to the entire network.
  • Registration Loss for Fissure Segmentation: The authors found that standard Dice loss for the auxiliary fissure segmentation task was unstable and difficult to optimize. To address this, they introduce a registration-based loss. This loss trains a registration model to align the predicted fissure with the ground truth. The idea is that a good fissure prediction should require minimal displacement to match the ground truth, thus penalizing large displacement fields.
  • Combined Learning Objective: The final loss function is a weighted sum of the Attentive Cross Entropy loss for lobe segmentation, the standard Dice loss for lobe segmentation, and the proposed registration loss for fissure segmentation.

Datasets

  • Private STLB Dataset: Used for training, validation, and testing. A subset of 207 cases was used, with 123 for training, 42 for validation, and 42 for testing. This dataset includes samples with pulmonary nodules, tuberculosis, and pneumonia, which can affect fissure visibility.
  • Public LUNA16 Dataset: A subset of 51 cases was used, with 41 for training and 10 for testing. This dataset is known to contain a large number of samples with pulmonary nodules.

Both datasets were preprocessed by resampling to 1mm x 1mm x 1mm spacing and normalizing voxel values within a [-1000, 400] HU window to [0, 1].

Key Insights and Ideas

  • Focusing on the "Hard" Parts: The core idea of the Attentive Cross Entropy loss is that accurately segmenting the difficult fissure regions is key to overall lobe segmentation. By explicitly weighting these regions higher, the model is guided to learn more robust representations for these challenging boundaries.
  • Integrated Fissure Prediction: The end-to-end fissure generation module is a clever way to leverage the lobe segmentation network's features for fissure prediction without adding significant architectural complexity or requiring a separate, potentially poorly performing, network branch. This allows the fissure prediction to benefit from the entire network's learned representations.
  • Registration as a Fissure Constraint: The shift from Dice loss to registration loss for fissure segmentation highlights a key insight: for thin, complex structures like fissures, geometric alignment and shape consistency are more important than simple voxel overlap. Registration loss directly enforces this geometric similarity.
  • Robustness to Pathologies: The paper demonstrates that their method is relatively insensitive to common lung pathologies like nodules, tuberculosis, and pneumonia, which is a significant advantage for real-world clinical applications.

Key Results

  • State-of-the-Art Performance: The proposed method achieved a mean Dice score of 97.83% on the private STLB dataset and 94.75% on the public LUNA16 dataset, outperforming several mainstream methods.
  • Improved Fissure Segmentation: Ablation studies showed that incorporating the Attentive Cross Entropy loss and the registration-based fissure loss significantly improved Dice scores and average symmetric surface distances (ASSD), particularly for challenging fissures like the right middle lobe and the small right horizontal fissure.
  • Qualitative Improvements: Visualizations show that the proposed method produces smoother lobe boundaries that better align with the actual fissure locations compared to other methods.

Limitations and Future Work

  • Hyperparameter Tuning: The paper mentions that $\lambda_1, \lambda_2, \lambda_3$ are hyper-parameters controlling the weights of different loss components. While they fix $\lambda_1$ and $\lambda_2$ to 1 and linearly increase $\lambda_3$ from 0 to 1, further exploration of optimal hyperparameter settings might be beneficial.
  • Computational Cost: While not explicitly detailed as a limitation, training deep learning models for 3D medical image segmentation is computationally intensive. The use of a single NVIDIA RTX 3090 GPU for training suggests significant computational resources are required.
  • Future Work: The authors suggest that their method's robustness to pathologies indicates its potential for clinical application. Further validation on larger and more diverse clinical datasets would be a natural next step. Exploring more advanced registration techniques or alternative fissure generation mechanisms could also be avenues for future research.

Summary updated

BibTeX
@inproceedings{zhang2023automatic,
  author = {Xiaofan Zhang and Qi Su and Na Wang and Jiawen Xie and Yinan Chen and Xiaofan Zhang},
  title = {{Automatic Lobe Segmentation Using Attentive Cross Entropy and End-to-End Fissure Generation}},
  booktitle = {ISBI},
  year = {2023},
  doi = {10.1109/isbi53787.2023.10230800},
}
Figure