Segmentation-Free PVC for Cardiac SPECT Using a Densely-Connected Multi-Dimensional Dynamic Network

Huidong Xie, Zhao Liu, Luyao Shi, Kathleen Greco, Xiongchao Chen, Bo Zhou, Attila Feher, John C. Stendahl, Nabil Boutagy, Tassos C. Kyriakides, Ge Wang, Albert J. Sinusas, Chi Liu

This paper, recently published in IEEE Transactions on Medical Imaging, presents a significant step forward in making cardiac nuclear imaging more accurate and less labor-intensive. Researchers from Yale University have developed a deep learning approach that corrects image blurring in SPECT scans without requiring the tedious manual "segmentation" of organs that has long been a bottleneck in the field.

TL;DR

Researchers have developed a "segmentation-free" deep learning model that corrects for partial volume effects in cardiac SPECT imaging. By using a novel densely-connected dynamic convolution mechanism and a clinically-informed loss function, the model achieves quantitative accuracy comparable to the manual "gold standard" while eliminating hours of human labor.

What Problem Does This Paper Address?

Single-photon emission computed tomography (SPECT) is a vital tool for measuring blood flow and tracer activity in the heart. However, it suffers from Partial Volume Effects (PVE)—a type of blurring caused by the system's limited spatial resolution. In cardiac imaging, this causes "spill-over," where signals from the heart muscle (myocardium) and the blood pool mix, leading to inaccurate measurements.

To fix this, scientists use Partial Volume Correction (PVC). The current gold standard, the "Iterative Yang" (iY) method, requires high-resolution anatomical maps from CT or MRI. The catch? A human expert must manually segment these anatomical images—a process that can take up to 10 hours per scan. Furthermore, any slight misalignment between the SPECT and CT scans can ruin the results. The authors sought a way to perform high-quality PVC automatically, using only the SPECT data itself.

Methods

The authors propose a specialized neural network based on a U-Net architecture, but with two major innovations:

  1. Densely-connected Multi-dimensional Dynamic (DC-Dy) Mechanism: Standard convolutional neural networks use "static" kernels—the filters are fixed once training is over. In this model, the kernels are "dynamic," meaning they adapt their weights based on the specific features of the input image. Borrowing from "Dense-net" logic, the model reuses feature maps from all preceding layers to calculate these dynamic weights, allowing the network to capture more complex contextual information without a massive increase in parameters.
  2. Intramyocardial Blood Volume (IMBV) Loss: Most AI models are trained to minimize simple pixel differences (like Mean Absolute Error). However, these don't always reflect clinical reality. The authors introduced a custom loss function based on IMBV—the ratio of activity in the myocardium to the blood pool. By forcing the AI to optimize for this specific clinical metric, the resulting images became much more quantitatively accurate.

Datasets

The study utilized 28 canine studies acquired on a GE Discovery NM/CT 570c dedicated cardiac scanner.

  • Tracer: Technetium-99m-labeled red blood cells (RBCs).
  • Ground Truth: The "gold standard" iY-PVC method (which uses manual segmentation and anatomical CT data).
  • Data Split: 15 studies for training (augmented to 1,020 volumes via rotation), 3 for validation, and 10 for testing.

Key Insights and Ideas

  • Adaptivity is Key: The "aha moment" here is that fixed filters aren't enough for the variability of medical imaging. By allowing the network to adjust its own filters for every patient (the dynamic mechanism), it can better handle the unique noise and blurring patterns of SPECT.
  • Clinical Knowledge > Generic Metrics: The inclusion of the IMBV loss function was crucial. While standard metrics like SSIM (Structural Similarity) improved slightly, the clinical measurement of blood volume saw a massive jump in accuracy when the model was specifically told to care about it.
  • Breaking the CT Dependency: Perhaps the most important insight is that the model performed nearly as well without anatomical CT data as it did with it. This suggests that the "missing" anatomical information is often implicitly present in the blurred SPECT data, and a sufficiently smart network can recover it.

Key Results

  • Comparable to Gold Standard: The proposed model (DC-Dy-U-BV) achieved an IMBV measurement of 0.169 ± 0.034, which is statistically indistinguishable from the manual iY-PVC reference of 0.167 ± 0.033.
  • Outperforming Standard AI: The dynamic mechanism significantly outperformed a standard U-Net (which had an IMBV of 0.193).
  • Robustness: The model proved to be more robust than the manual method in cases where the SPECT and CT scans were slightly misaligned—a common real-world problem that usually causes the manual method to fail.

Limitations and Future Work

While promising, the study was limited to canine data. The physiological differences in humans (size, heart rate, and attenuation) mean the model will need to be retrained and validated on human cohorts. Additionally, the study focused on a specific tracer (labeled RBCs); future work is needed to see if this "segmentation-free" approach holds up for other common tracers used in myocardial perfusion imaging. Finally, the authors suggest combining this PVC method with corrections for "limited-view" artifacts common in stationary SPECT scanners.

Summary updated

Abstract

In nuclear imaging, limited resolution causes partial volume effects (PVEs) that affect image sharpness and quantitative accuracy. Partial volume correction (PVC) methods incorporating high-resolution anatomical information from CT or MRI have been demonstrated to be effective. However, such anatomical-guided methods typically require tedious image registration and segmentation steps. Accurately segmented organ templates are also hard to obtain, particularly in cardiac SPECT imaging, due to the lack of hybrid SPECT/CT scanners with high-end CT and associated motion artifacts. Slight mis-registration/mis-segmentation would result in severe degradation in image quality after PVC. In this work, we develop a deep-learning-based method for fast cardiac SPECT PVC without anatomical information and associated organ segmentation. The proposed network involves a densely-connected multi-dimensional dynamic mechanism, allowing the convolutional kernels to be adapted based on the input images, even after the network is fully trained. Intramyocardial blood volume (IMBV) is introduced as an additional clinical-relevant loss function for network optimization. The proposed network demonstrated promising performance on 28 canine studies acquired on a GE Discovery NM/CT 570c dedicated cardiac SPECT scanner with a 64-slice CT using Technetium-99m-labeled red blood cells. This work showed that the proposed network with densely-connected dynamic mechanism produced superior results compared with the same network without such mechanism. Results also showed that the proposed network without anatomical information could produce images with statistically comparable IMBV measurements to the images generated by anatomical-guided PVC methods, which could be helpful in clinical translation.
BibTeX
@article{xie2023segmentation,
  author = {Huidong Xie and Zhao Liu and Luyao Shi and Kathleen Greco and Xiongchao Chen and Bo Zhou and Attila Feher and John C. Stendahl and Nabil Boutagy and Tassos C. Kyriakides and Ge Wang and Albert J. Sinusas and Chi Liu},
  title = {{Segmentation-Free PVC for Cardiac SPECT Using a Densely-Connected Multi-Dimensional Dynamic Network}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2023},
  url = {https://openreview.net/forum?id=_sILPBnx35m},
}

Loading related content...

Figure