Computed tomography-based deep-learning prediction of lymph node metastasis risk in locally advanced gastric cancer

An-qi Zhang, Huiping Zhao, Fei Li, Pan Liang, Jianbo Gao, Ming ChengFrontiers in Oncology 2022 · DOI

TL;DR

Researchers have developed a non-invasive deep-learning pipeline that uses preoperative CT scans to predict lymph node metastasis (LNM) in patients with locally advanced gastric cancer. By leveraging transfer learning with a ResNet-SVM architecture, the model significantly outperformed both traditional clinical assessments and handcrafted radiomics, providing a more accurate tool for personalized surgical planning.

What Problem Does This Paper Address?

In locally advanced gastric cancer (LAGC), knowing whether the cancer has spread to the lymph nodes is critical for deciding between different surgical approaches or the need for neoadjuvant therapy. Currently, radiologists rely on visual inspection of Computed Tomography (CT) images, but this method is notoriously unreliable, with accuracy rates hovering between 50% and 70%.

While "radiomics"—the extraction of hundreds of handcrafted geometric and textural features from images—has shown promise, it is labor-intensive and highly subjective, as it requires precise manual contouring of tumor boundaries. The authors sought to determine if deep learning (DL) could provide a more automated, objective, and accurate prediction of LNM risk.

Methods

The study employed a retrospective design using a "transfer learning" strategy. Because medical datasets are often too small to train complex neural networks from scratch, the researchers used five established convolutional neural networks (CNNs) pre-trained on the massive ImageNet dataset: ResNet50, VGG16, VGG19, Xception, and InceptionV3.

The workflow involved:

  1. Feature Extraction: Using the CNNs as fixed feature extractors on the largest cross-sectional area of the tumor from CT scans.
  2. Harmonization: Applying the "ComBat" technique to remove "batch effects" caused by different CT scanners or imaging protocols.
  3. Feature Selection: Using the LASSO (Least Absolute Shrinkage and Selection Operator) algorithm to identify the most predictive features.
  4. Classification: Training a Support Vector Machine (SVM) to classify the risk of LNM based on the selected features.
  5. Visualization: Using Grad-CAM (Gradient-weighted Class Activation Mapping) to produce heatmaps, allowing researchers to see which parts of the image the model prioritized.

Datasets

The study used a combined dataset of 523 patients with pathologically confirmed LAGC, treated at the First Affiliated Hospital of Zhengzhou University between 2012 and 2019.

  • Training Cohort: 367 patients.
  • Testing Cohort: 156 patients. The dataset included clinical factors (age, sex, tumor size, T-stage) and contrast-enhanced CT scans.

Key Insights and Ideas

  • Deep Learning vs. Radiomics: The abstract, high-level features extracted by CNNs proved superior to handcrafted radiomics. This suggests that the "hidden" patterns captured by deep layers of a neural network are more representative of metastatic potential than human-defined shapes or textures.
  • Efficiency Gains: Unlike radiomics, which requires meticulous manual segmentation of the entire tumor volume, the DL approach only requires a simple bounding box around the tumor. This makes it far more viable for a high-throughput clinical environment.
  • The Power of ResNet: Among the five architectures tested, ResNet50 emerged as the clear winner. The authors attribute this to its residual learning framework, which effectively captures complex hierarchical features without the degradation issues found in deeper VGG-style networks.
  • Diminishing Returns of Fusion: Interestingly, combining deep learning features with clinical data or radiomics did not significantly boost performance. This implies that the deep learning features are robust enough to capture the relevant predictive information already present in clinical indicators like tumor size.

Key Results

The ResNet-SVM model was the top performer across the board in the testing cohort:

  • AUC (Area Under the Curve): 0.796, significantly higher than the radiomics model (0.704) and the clinical signature (0.683).
  • Accuracy: 75.2% (compared to 61.8% for radiomics).
  • Sensitivity: 80.2%, which is particularly important in a clinical setting to ensure that metastatic cases are not missed.
  • Qualitative Findings: Grad-CAM visualizations confirmed that the model focused primarily on the tumoral lesion and the perifocal (surrounding) areas, which are biologically relevant zones for cancer spread.

Limitations and Future Work

The authors acknowledge several hurdles before this can be used in routine practice:

  • Single-Center Study: The data comes from one hospital. To prove the model's generalizability, it needs to be validated against diverse datasets from different geographic regions and populations.
  • Manual Slice Selection: While the model doesn't need full 3D segmentation, a radiologist still has to manually select the "best" 2D slice and draw a bounding box. Future work could involve fully automating the detection and selection process.
  • Data Volume: While transfer learning is effective, the authors note that a larger, multi-center annotated dataset would allow for "fine-tuning" the networks specifically for gastric tissue, potentially pushing the AUC even higher.

Summary updated

Abstract

The novel and noninvasive deep learning approach could provide efficient and accurate prediction of lymph node metastasis in LAGC, and benefit clinical decision making of therapeutic strategy.
BibTeX
@inproceedings{zhang2022computed,
  author = {An-qi Zhang and Huiping Zhao and Fei Li and Pan Liang and Jianbo Gao and Ming Cheng},
  title = {{Computed tomography-based deep-learning prediction of lymph node metastasis risk in locally advanced gastric cancer}},
  booktitle = {Frontiers in Oncology},
  year = {2022},
  doi = {10.3389/fonc.2022.969707},
}
Figure