The Problem with Existing Segmentation Accurate segmentation of lung structures from CT images is foundational for diagnosis, radiotherapy planning, and radiomics feature extraction. Existing deep learning methods often struggle with boundary precision and fail to capture the long-range spatial relationships between anatomical structures that are critical for reliable lung CT analysis.
The GEANet Framework This study introduced GEANet (Graph Edge Attention Network), a novel segmentation architecture that combines Graph Neural Networks (GNN) with radiomics features. By modeling the lung CT image as a graph where nodes represent image regions and edges encode spatial relationships, GEANet can capture structural context that standard convolutional networks miss.
Key Innovations The framework incorporates three components: a GNN backbone for relational reasoning, a radiomics feature extractor using PyRadiomics for quantitative texture and shape descriptors, and a boundary refinement module to sharpen segmentation edges. Together, these address both global structure and fine-grained detail.
Performance Achievement GEANet achieved an Intersection over Union (IoU) of 98.07% on the test set, outperforming state-of-the-art segmentation models. This level of precision is clinically significant because even small segmentation errors in lung CT can propagate into downstream radiomics measurements and treatment planning volumes.
Segmentation as a Foundation Lung CT segmentation is not an end in itself - it is the prerequisite for almost every quantitative imaging analysis. Radiomics pipelines, tumor volume measurements, lymph node assessment, and radiotherapy target delineation all depend on the quality of the initial segmentation. Errors in segmentation compound into errors in all downstream tasks.
Challenges of Manual Segmentation Manual segmentation by radiologists is the current clinical standard but is time-consuming, subject to inter-observer variability, and impractical at scale. Studies have shown that the same lung structure can vary by 10-20% in volume depending on who segments it - a difference that can meaningfully change treatment decisions.
Limitations of Standard CNN Approaches Standard convolutional neural networks like U-Net process images locally, capturing features within a fixed receptive field. This limits their ability to model long-range dependencies between distant anatomical structures. In lung CT, where the relationship between the lobes, vessels, and airways matters for accurate delineation, local feature extraction alone is insufficient.
Radiomics Integration Rationale Adding radiomics features from PyRadiomics provides hand-crafted quantitative descriptors - including first-order statistics, texture features (GLCM, GLRLM), and shape descriptors - that encode domain knowledge not always captured by learned features. This hybrid approach combines the flexibility of deep learning with the interpretability of established radiomics.
Graph Representation of CT Images The CT image was converted to a graph by treating image superpixels or regions as nodes. Edges were constructed based on spatial proximity and feature similarity. This graph representation allows the GNN to propagate information across distant regions, enabling the model to reason about global lung anatomy rather than just local texture.
GNN with Edge Attention GEANet uses an edge attention mechanism that assigns different weights to different edges based on their relevance. This allows the model to selectively focus on the most informative relationships - for example, emphasizing the boundary between lung parenchyma and the pleura while downweighting less discriminative connections deeper within the tissue.
Radiomics Feature Extraction PyRadiomics was used to extract quantitative features from segmented regions including first-order intensity statistics, GLCM texture features, and shape descriptors. These features were concatenated with learned GNN features and fed into the final segmentation head, enriching the model's representation with domain-specific quantitative descriptors.
Boundary Refinement Module A dedicated post-processing module was applied to sharpen predicted segmentation boundaries. This module used gradient-based techniques to detect and correct boundary-region misclassifications - the most common failure mode of graph-based segmentation methods where edge regions have ambiguous class membership.
Dataset Composition The model was trained and evaluated on a curated lung CT dataset with expert-annotated segmentation masks. The dataset included diverse cases representing different lung pathologies, patient demographics, and CT acquisition parameters to ensure the model would generalize across clinical variability.
Data Augmentation To prevent overfitting, extensive data augmentation was applied including random rotation, flipping, elastic deformation, and intensity perturbation. Elastic deformation specifically simulates the natural variation in lung shape across breathing cycles and patient positions, which is a major source of real-world variability.
Training Protocol The model was trained using a combined loss function incorporating Dice loss for region overlap and cross-entropy loss for pixel-wise classification. The Adam optimizer was used with learning rate scheduling. The boundary refinement module was trained end-to-end with the main network rather than as a separate post-processing step.
Comparison Methods GEANet was benchmarked against established segmentation models including U-Net, DeepLab, and prior graph-based approaches. All comparison models were retrained on the same dataset under identical conditions to ensure a fair evaluation. IoU, Dice score, and Hausdorff distance were used as complementary metrics.
Primary Result GEANet achieved an IoU of 98.07% on the held-out test set, representing a meaningful improvement over the best-performing comparison model. An IoU above 98% is approaching the level of inter-expert agreement among radiologists, suggesting the model has reached near-human segmentation quality.
Dice Score and Hausdorff Distance Corresponding Dice scores confirmed the IoU result, and Hausdorff distance measurements showed that GEANet produced significantly tighter boundary predictions than comparison models. The Hausdorff metric is particularly informative because it captures worst-case boundary deviations, which are clinically relevant for radiotherapy planning.
Ablation Study Results An ablation study demonstrated that each component contributed meaningfully to performance. Removing the GNN and relying only on CNN features reduced IoU by several percentage points. Removing radiomics features caused a smaller but still significant drop. The boundary refinement module had the largest single impact on Hausdorff distance.
Qualitative Analysis Visual inspection of segmentation outputs showed that GEANet correctly resolved challenging cases including juxtapleural nodules adjacent to the chest wall, vessels overlapping with lobe boundaries, and cases with pleural effusion that distorts normal lung anatomy - failure cases for standard CNN models.
Computational Overhead Graph construction and GNN inference are computationally more expensive than standard CNN forward passes. The added latency may be acceptable for offline radiotherapy planning but could limit real-time applications during interventional procedures. Hardware acceleration and graph construction optimization are areas for future engineering work.
Single-Institution Dataset The model was trained and evaluated on data from a single institution. Performance may degrade when applied to CT images acquired with different scanner manufacturers, reconstruction kernels, or clinical protocols. Multi-center prospective evaluation is needed to assess generalizability.
Organ Generalization GEANet was developed specifically for lung CT segmentation. Extending the framework to other thoracic structures (heart, esophagus, spinal cord) relevant to radiotherapy planning would require retraining and validation but is architecturally feasible given the graph-based design.
Future Directions Future work should evaluate GEANet on downstream tasks - specifically, whether better segmentation translates into more accurate radiomics features and improved predictive models for treatment outcomes. Integration into clinical radiotherapy planning software and prospective testing in workflow studies would establish real-world utility.