Enhancing Early Lung Cancer Diagnosis: Predicting Lung Nodule Progression in Follow-Up Low-Dose CT Scan with Deep Generative Model

Yifan Wang, Chuan Zhou, Lei Ying, Heang‐Ping Chan, Elizabeth Lee, Aamer Chughtai, Lubomir M. Hadjiiski, Ella A. KazerooniCancers 2024 · DOI

TL;DR

This paper introduces a novel deep generative model (GP-WGAN) that can predict the future appearance of lung nodules in low-dose CT scans using only the initial scan. This "virtual" follow-up imaging significantly improves the accuracy of lung cancer risk assessment at the baseline screening, potentially leading to earlier diagnosis and treatment.

What Problem Does This Paper Address?

Early detection of lung cancer dramatically improves patient survival rates. Low-dose computed tomography (LDCT) screening is effective at identifying lung cancer early, but it also frequently detects indeterminate nodules. Current practice requires patients to wait for follow-up scans to assess nodule stability and growth, which can lead to diagnostic delays and increased patient anxiety. Existing predictive models often rely on manually engineered features or require prior scans, limiting their utility. This paper aims to bridge this gap by developing a method to predict future nodule appearance from a single baseline scan, enabling more immediate and accurate risk stratification.

Methods

The core of this work is the Growth Predictive Wasserstein Generative Adversarial Network (GP-WGAN). This model takes a baseline LDCT nodule image as input and generates a synthetic image representing how that nodule might look in a follow-up scan (specifically, one year later).

The GP-WGAN consists of two main components:

  • Predictor Network: Implemented using a modified U-Net architecture, this network is responsible for generating the "virtual" follow-up nodule images.
  • Discriminator Network: A deep convolutional neural network (DCNN) that tries to distinguish between real follow-up nodule images and the synthetic ones generated by the predictor.

The training process is adversarial: the predictor tries to fool the discriminator, while the discriminator gets better at detecting fakes. This competition drives the predictor to generate increasingly realistic images. A novel generative loss function was designed, combining:

  • L1 Loss: For pixel-wise accuracy.
  • Structural Similarity Index (SSIM) Loss: To preserve structural integrity.
  • Learned Perceptual Loss (LLP): Using a pre-trained ResNet-18 model to capture high-level visual quality and semantic similarity.
  • Adversarial Loss: To ensure realistic outputs.

Once trained, only the predictor network is used to generate GP-nodules. These GP-nodules are then fed into a pre-existing Lung Cancer Risk Prediction (LCRP) model, which was previously developed using serial radiomics. The performance of the LCRP model using GP-nodules is then compared against using real baseline nodules and real follow-up nodules.

Datasets

The study utilized data from the National Lung Screening Trial (NLST).

  • Training/Validation Set: Comprised of 776 subjects, including 1121 pairs of nodule images from baseline (T0) and follow-up (T1 or T2) scans.
  • Test Set: An independent set of 450 nodules from 450 subjects, with baseline (T0) and 1-year follow-up (T1) scans.

The dataset included subjects diagnosed with lung cancer and those with benign nodules. The data was preprocessed by resampling to isotropic volumes and extracting 32mm x 32mm regions of interest (ROIs) around manually marked nodule centers.

Key Insights and Ideas

  • "Virtual" Follow-up is Feasible: The most significant insight is that a deep generative model can effectively predict the appearance of lung nodules in future CT scans using only the initial scan. This bypasses the need for actual follow-up imaging for initial risk assessment.
  • Generative Models Enhance Risk Prediction: By providing these "virtual" follow-up images, the GP-WGAN significantly improves the performance of existing lung cancer risk prediction models, particularly in stratifying risk at the baseline screening.
  • Beyond Simple Growth Models: The use of a multi-component loss function (L1, SSIM, LLP, Adversarial) is crucial for generating realistic and structurally sound nodule images, capturing complex growth patterns that simpler models might miss.
  • Explainable AI Potential: The generated GP-nodules offer a step towards more explainable AI in medical imaging, potentially allowing radiologists to better understand the basis of risk predictions.

Key Results

  • Comparable Performance to Real Follow-up: The LCRP model using GP-nodules achieved a test AUC of 0.827 ± 0.028, which was comparable to the AUC of 0.862 ± 0.028 achieved by the same LCRP model using real follow-up nodule images (p = 0.071).
  • Outperforms Baseline Models: The LCRP model with GP-nodules significantly outperformed baseline methods like Lung-RADS and the Brock model when using only baseline scans for risk prediction.
  • Improved Risk Stratification: Net Reclassification Index (NRI) analysis showed that the LCRP model using GP-nodules significantly improved risk stratification compared to Lung-RADS (NRI = 0.38, p < 0.001), the Brock model (NRI = 0.20, p = 0.03), and the LCRP model with real baseline nodules (NRI = 0.20, p = 0.004).
  • Specific Nodule Types Benefit: The GP-nodules showed particular benefit in classifying nodules that were solid, spiculated, or within the 6-14 mm diameter range.

Limitations and Future Work

  • Limited Training Data for Malignant Nodules: The study acknowledges that the relatively small training sample size of malignant nodules, with their wide range of progression paces, might limit the GP-WGAN's capability in predicting rapid growth rates. Figure A4 provides examples where the model underestimated growth.
  • External Validation Needed: The authors suggest that extensive external validation and reader studies are necessary to further assess the model's performance and clinical utility.
  • Further Refinement of Loss Functions: While the current loss function is effective, ongoing research may explore even more sophisticated metrics for image similarity.
  • Broader Clinical Integration: Future work could explore integrating GP-nodules into other AI-driven tools for nodule segmentation and temporal analysis, and ultimately, into clinical workflows to accelerate diagnosis and patient management.

Summary updated

Abstract

Early diagnosis of lung cancer can significantly improve patient outcomes. We developed a Growth Predictive model based on the Wasserstein Generative Adversarial Network framework (GP-WGAN) to predict the nodule growth patterns in the follow-up LDCT scans. The GP-WGAN was trained with a training set (N = 776) containing 1121 pairs of nodule images with about 1-year intervals and deployed to an independent test set of 450 nodules on baseline LDCT scans to predict nodule images (GP-nodules) in their 1-year follow-up scans. The 450 GP-nodules were finally classified as malignant or benign by a lung cancer risk prediction (LCRP) model, achieving a test AUC of 0.827 ± 0.028, which was comparable to the AUC of 0.862 ± 0.028 achieved by the same LCRP model classifying real follow-up nodule images (<i>p</i> = 0.071). The net reclassification index yielded consistent outcomes (NRI = 0.04; <i>p</i> = 0.62). Other baseline methods, including Lung-RADS and the Brock model, achieved significantly lower performance (<i>p</i> < 0.05). The results demonstrated that the GP-nodules predicted by our GP-WGAN model achieved comparable performance with the nodules in the real follow-up scans for lung cancer diagnosis, indicating the potential to detect lung cancer earlier when coupled with accelerated clinical management versus the current approach of waiting until the next screening exam.
BibTeX
@inproceedings{wang2024enhancing,
  author = {Yifan Wang and Chuan Zhou and Lei Ying and Heang‐Ping Chan and Elizabeth Lee and Aamer Chughtai and Lubomir M. Hadjiiski and Ella A. Kazerooni},
  title = {{Enhancing Early Lung Cancer Diagnosis: Predicting Lung Nodule Progression in Follow-Up Low-Dose CT Scan with Deep Generative Model}},
  booktitle = {Cancers},
  year = {2024},
  doi = {10.3390/cancers16122229},
}
Figure