Abstract
Functional avoidance radiation therapy for lung cancer patients aims to limit dose delivery to highly functional lung. However, clinical functional imaging suffers from many shortcomings, including the need for exogenous contrasts and longer processing times. In this study, we present a new approach to derive lung functional images using a deep convolutional neural network (CNN) to learn and exploit the underlying functional information in CT images and generate functional perfusion images. 99mTc MAA SPECT/CT scans of 30 lung cancer patients were retrospectively analyzed. The CNN model was trained using a randomly selected dataset of 25 patients and tested using the remaining 5 subjects. Our study showed that it is feasible to derive perfusion images from CT images. Using the deep neural network with discrete labels, the main defect regions can be predicted. This technique holds promise for providing lung function images for image-guided functional lung avoidance radiation therapy.
Introduction
In lung cancer treatment, radiation therapy is a double-edged sword: it must deliver a high dose to the tumor while sparing healthy, functional lung tissue to avoid injuries like pneumonitis or fibrosis. The current gold standard for identifying functional lung regions (perfusion) is the SPECT/CT scan using a radioactive tracer ($^{99m}$Tc MAA). However, SPECT scans are expensive, invasive, and not always available in radiation oncology departments.
While CT scans are standard for treatment planning, they are primarily anatomical. This study explores whether a deep learning model can bridge this gap by "seeing" functional information hidden within the Hounsfield Units (HU) of a standard CT image. HU values represent the air-to-tissue ratio, which is intrinsically linked to lung function and underlying vascular structure.
Methods
The researchers employed a 3D U-Net architecture, a standard in medical image synthesis. The model consists of a contraction path to capture broad lung context and an expansion path for precise localization of blood flow defects.
The core experiment compared two data-labeling strategies to determine the most effective way to train the model:
- Scenario 1 (Data Discretization): Continuous SPECT values were converted into 11 discrete intervals.
- Scenario 2 (Data Reduction): The model focused only on "low-function" regions (values below 0.5) to predict defects.
The study used a retrospective dataset of 30 lung cancer patients. CT and SPECT images were co-registered, masked to include only lung parenchyma, and resized to a $128 \times 128 \times 64$ matrix. 25 patients were used for training and 5 for testing.
Key Results and Insights
The study demonstrated that synthetic perfusion mapping is feasible. A significant finding was that Scenario 1 (Discretization) outperformed Scenario 2 (Reduction). Scenario 1 achieved an average Spearman correlation coefficient ($r$) of $0.53 \pm 0.14$, while Scenario 2 lagged at $0.39 \pm 0.14$. This suggests that the network performs better when trained on the full range of lung function data, as healthy tissue provides necessary context for identifying deficits.
Qualitatively, visual analysis showed the model was particularly effective at identifying major perfusion defects in the upper lobes. Furthermore, once trained, the CNN can generate these functional maps almost instantaneously, offering a significant speed advantage over traditional SPECT processing.
Limitations and Future Directions
While promising, this preliminary study has limitations. The sample size of 30 patients is small for deep learning, and a correlation of 0.53 is not yet high enough to replace SPECT in clinical decision-making. Future work will involve larger datasets and the potential integration of 4DCT (capturing lung motion) to provide more functional clues to the neural network. This technique holds the promise of providing non-invasive lung function images for image-guided radiation therapy.
Summary updated