Abstract
Accurate anatomical mapping of pulmonary segments is critical for modern lung-sparing surgeries, such as segmentectomy. However, traditional voxel-based deep learning models often struggle with the high memory demands and resolution constraints required to capture the intricate, non-visible boundaries of these segments. This paper introduces ImPulSe (Implicit Pulmonary Segment), a novel framework that utilizes implicit surface functions to represent lung anatomy as a continuous field. By leveraging multi-scale feature pyramids and a lightweight MLP decoder, ImPulSe achieves state-of-the-art reconstruction accuracy while significantly reducing computational overhead. Our findings highlight the importance of bronchial and arterial structures as anatomical landmarks and reveal a counter-intuitive "lobe paradox" where the inclusion of lobe-level data can degrade segmentation performance.
Introduction
In the treatment of early-stage lung cancer, segmentectomy—the surgical removal of a specific bronchopulmonary segment—has emerged as a preferred alternative to total lobectomy. By preserving healthy lung tissue, segmentectomy maintains better post-operative pulmonary function and accelerates patient recovery. Unlike lung lobes, which are separated by visible physical fissures, the 18 pulmonary segments lack distinct anatomical boundaries. Instead, they are defined functionally by the specific branches of the bronchi and pulmonary arteries that supply them.
For surgeons, the precise localization of these segments is a complex 3D puzzle. Traditional deep learning approaches, primarily based on U-Net architectures, treat this as a voxel-wise classification task. However, these methods are limited by the "curse of dimensionality": high-resolution grids require massive GPU memory, while downsampled grids lose the fine-grained details of the small vessels that define segmental borders. To address these limitations, we present ImPulSe, a model that shifts the paradigm from discrete voxel grids to continuous implicit representations.
The ImPulSe Framework
Implicit Surface Representation
ImPulSe moves away from classifying pixels in a fixed grid. Instead, it learns a continuous mathematical function of 3D space. The core idea is to represent the lung volume as a field where any coordinate $(x, y, z)$ can be queried to determine its segmental identity.
Architecture
The model consists of two primary components:
- Feature Pyramid Encoder: A 3D ResNet18 backbone processes the input CT data to extract multi-scale digital signatures. This "feature pyramid" captures both global anatomical context and local structural details.
- Implicit Decoder: Rather than a heavy convolutional decoder, ImPulSe employs a lightweight two-layer Multi-Layer Perceptron (MLP). To determine the segment of a specific point, the model performs tri-linear interpolation on the feature pyramid at that coordinate. These interpolated features, combined with the raw spatial coordinates, are fed into the MLP to produce a classification.
Computational Efficiency
Because the model operates on a continuous field, it does not need to process every voxel during the training phase. By sampling random points in space, ImPulSe reduces training time by over 50% compared to standard voxel-based networks. Furthermore, during inference, the model is resolution-independent; it can generate reconstructions at any density, providing smooth, anatomically plausible surfaces without the "blocky" artifacts typical of traditional methods.
Experimental Results
Dataset and Benchmarking
The study utilized a multi-center dataset of 800 CT scans, meticulously annotated by experts. These annotations included the 18 pulmonary segments as well as the underlying bronchi, arteries, and veins. The model was evaluated against established architectures, including U-Net and DeepLabv3.
ImPulSe achieved a Dice score of 84.63%, outperforming U-Net (83.25%) and DeepLabv3 (82.44%). Notably, the model achieved these results with 14% fewer parameters, demonstrating a superior balance of performance and efficiency.
The Importance of Vascular Landmarks
A key contribution of this research is the ablation study on input modalities. The researchers tested various combinations of anatomical inputs, including raw images, lobes, bronchi, arteries, and veins. The results confirmed that Bronchi and Arteries (BAV) are the most critical inputs for accurate reconstruction. Since segments are functionally defined by these structures, providing them as explicit inputs allows the model to map boundaries with much higher confidence.
The Lobe Paradox
One of the most surprising findings was the impact of lobe information. Intuitively, knowing which lobe a point belongs to should help narrow down the possible segments. However, adding lobe data as an input resulted in a 9% drop in Dice score. The authors hypothesize that the relatively simple geometry of the lung lobes may cause the model to overfit to these coarse boundaries, leading it to ignore the complex internal branching of the bronchi and vessels that actually dictate segmental limits.
Discussion and Future Work
While ImPulSe represents a significant step forward, challenges remain in the modeling of pulmonary veins. While intersegmental veins are the true anatomical "fences" between segments, they are notoriously difficult to segment accurately due to their variability and thin structures. Currently, these are excluded from primary evaluation metrics to maintain consistency, but future iterations of ImPulSe could benefit from more sophisticated venous modeling.
Additionally, the authors suggest that the framework could be enhanced by incorporating "template deformation." By warping a standard anatomical template to fit the specific geometry of a patient's lung, the model could achieve even greater robustness in cases with atypical anatomy or pathological distortions.
Conclusion
ImPulSe demonstrates that implicit neural representations are highly effective for the complex task of pulmonary segment reconstruction. By focusing on the underlying functional anatomy (bronchi and arteries) and utilizing a continuous coordinate-based approach, the model provides a high-fidelity, memory-efficient tool for surgical planning, potentially improving outcomes for patients undergoing lung-sparing procedures.
Summary updated