AnatomicalNets: A Multi-Structure Segmentation and Contour-Based Distance Estimation Pipeline for Clinically Grounded Lung Cancer T-Staging

1. Problem Statement

Accurate T-staging of lung cancer is critical for prognosis and treatment planning. According to the International Association for the Study of Lung Cancer (IASLC) guidelines, T-staging is not merely a function of tumor size but also depends on:

• Quantitative dimensions: Tumor size thresholds (e.g., T1 ≤3cm, T2 >3–5cm, etc.).
• Anatomical context: Proximity to or invasion of specific structures (mediastinum, diaphragm, chest wall, carina, etc.).

Current Limitations:
Most existing deep learning approaches treat T-staging as a standard image classification problem (end-to-end CNNs). These "black-box" models fail to:

1. Explicitly model the spatial relationships and anatomical context required by clinical guidelines.
2. Provide interpretable reasoning for their decisions (e.g., why a tumor is T4).
3. Handle fine-grained distinctions where a small change in distance or size alters the stage.
4. Generalize well to borderline cases due to a lack of explicit rule adherence.

2. Methodology: The AnatomicalNets Pipeline

The authors propose AnatomicalNets, a multi-stage, hybrid pipeline that reformulates T-staging as a measurement and rule-based inference problem rather than a direct classification task. The pipeline consists of three main stages:

Stage 1: Multi-Structure Segmentation & Detection

Instead of a single classifier, the system employs three dedicated Encoder-Decoder networks (collectively called AnatomicalNets) and a detection model:

• LungNet: Segments lung parenchyma. Uses a DenseNet-121 encoder and a standard U-Net decoder.
• MediNet: Segments the mediastinum. Uses a DenseNet-121 encoder and a UNet++ decoder to improve feature aggregation.
• TumorNet: Segments the tumor. Operates on cropped Regions of Interest (ROIs). Uses a ResNet-152 encoder and a Feature Pyramid Network (FPN) decoder.
• Tumor Detection: A YOLOv11 model is trained to localize tumor ROIs in CT slices, which are then fed into TumorNet for precise segmentation.

Data Strategy: Since the primary dataset (Lung-PET-CT-Dx) lacks ground truth masks for all structures, the authors trained these networks on multiple secondary public datasets (e.g., MSD, SAROS, COVID-19 datasets) and applied them to the primary dataset as an external validation set to generate surrogate masks.

Stage 2: Diaphragm Estimation & Contour-Based Measurement

Since dedicated diaphragm segmentation datasets are scarce, the authors developed a heuristic estimation technique:

• The diaphragm boundary is estimated by analyzing the lowest 10% of the lung mask contour pixels.
• Quantitative Metrics Extraction: Using the generated segmentation masks, the system calculates:
  1. Maximum Tumor Dimension: The largest Euclidean distance between contour points (considering width, height, and depth via slice count).
  2. Distance to Structures: The minimum pixel distance between the tumor contour and the lung walls, mediastinum, and estimated diaphragm.
  3. Invasion Logic: A distance of zero implies direct invasion.

Stage 3: Deterministic Decision Module

The extracted quantitative features are fed into a rule-based decision engine that strictly adheres to IASLC guidelines:

• Size Check: Assigns T1–T4 based on diameter thresholds.
• Invasion Check: Overrides size-based staging if invasion of critical structures (mediastinum, diaphragm, etc.) is detected, immediately classifying the case as T4.

3. Key Contributions

1. Paradigm Shift: Moves away from end-to-end classification toward a clinically grounded, interpretable pipeline that mimics the radiologist's decision process (measure $\rightarrow$ compare $\rightarrow$ rule).
2. Specialized Architecture: Design of three distinct segmentation networks (LungNet, MediNet, TumorNet) with optimized encoder-decoder combinations (DenseNet, UNet++, ResNet, FPN) for specific anatomical structures.
3. Cross-Dataset Inference: A strategy to train on diverse public datasets and apply them to a primary dataset lacking full annotations, generating surrogate ground truth masks.
4. Heuristic Diaphragm Estimation: A novel, low-cost method to estimate diaphragm position without requiring a dedicated segmentation model or dataset.
5. Transparent Performance: Provides per-stage F1-scores and a confusion matrix, highlighting where the model succeeds and fails, unlike prior works that often report only aggregate accuracy.

4. Experimental Results

The system was evaluated on the Lung-PET-CT-Dx dataset (220 patients, 31,717 slices).

• Overall Classification Accuracy: 91.36%.
• Per-Stage F1-Scores:
  • T1: 0.93
  • T2: 0.89
  • T3: 0.96
  • T4: 0.90
• Comparison with Traditional CNNs:
  • The proposed method significantly outperformed standard CNN architectures (ResNet152, DenseNet121, Swin Transformer) which achieved accuracies between 37% and 40% and F1-scores around 39%.
• Segmentation Performance:
  • LungNet achieved 97.82% DSC.
  • MediNet achieved 93.39% DSC.
  • TumorNet achieved 89.68% DSC.
• Detection Performance: YOLOv11 achieved 94.3% mAP50 and 60.0% mAP50-95.

5. Significance and Limitations

Significance:

• Interpretability: The pipeline offers a "glass-box" approach where the decision logic is transparent and aligns with medical guidelines, increasing trust for clinical adoption.
• Clinical Relevance: By explicitly modeling anatomical distances and invasion, it addresses the specific spatial complexities of lung cancer staging that generic classifiers miss.
• Bottleneck Identification: The study identifies that the limitation in prior work lies in feature design (lack of anatomical context) rather than classifier capacity.

Limitations & Future Work:

• T4 Misclassification: The model struggles with T4 cases involving multifocal disease (separate nodules in different lobes) or invasion of structures not currently modeled (e.g., trachea, esophagus, great vessels).
• Diaphragm Heuristic: The diaphragm estimation is a heuristic based on lung contours and may fail in atypical anatomies.
• 2D vs. 3D: The current pipeline processes 2D slices, potentially missing volumetric tumor characteristics.
• Future Plans: The authors plan to curate true diaphragm segmentation datasets, extend the pipeline to full 3D volumetric analysis, and integrate lymph node and metastasis detection for complete TNM staging.

In conclusion, AnatomicalNets demonstrates that integrating deep learning segmentation with explicit clinical rules yields superior, interpretable, and clinically reliable results for lung cancer T-staging compared to traditional black-box classification methods.