Implementation of GAN-Based, Synthetic T2-Weighted Fat Saturated Images in the Routine Radiological Workflow Improves Spinal Pathology Detection

Sarah Schlaeger, Katharina Drummer, Malek El Husseini, Florian Kofler, Nico Sollmann, Severin Schramm, Claus Zimmer, Jan S. Kirschke, Benedikt WiestlerDiagnostics 2023 · DOI

TL;DR

This paper demonstrates that synthetic T2-weighted fat-saturated (T2-w fs) MRI images, generated by a Generative Adversarial Network (GAN), can significantly improve the detection and grading of spinal pathologies. By virtually adding these synthetic images to standard T1-w and non-fs T2-w scans, radiologists achieved higher diagnostic accuracy, mimicking the benefits of actual T2-w fs sequences without increasing scan time.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) is crucial for diagnosing spinal pathologies, and T2-weighted fat-saturated (T2-w fs) sequences are particularly valuable for enhancing contrast and visualizing abnormalities like edema and inflammation. However, in routine clinical practice, these T2-w fs sequences are often omitted due to time constraints or motion artifacts, leading to potentially missed diagnoses or the need for rescans. This creates a gap where valuable diagnostic information is lost, impacting patient care. The paper addresses this by exploring how Generative Adversarial Networks (GANs) can be used to create synthetic T2-w fs images from existing T1-w and non-fs T2-w scans, effectively bridging this diagnostic gap without altering the standard imaging protocol.

Methods

The core of this study lies in the implementation and evaluation of a GAN for generating synthetic T2-w fs MRI images. The GAN architecture is based on the pix2pix framework, utilizing a U-Net encoder-decoder for the generator and a patch-based discriminator. The generator is trained to produce realistic T2-w fs images from T1-w and non-fs T2-w inputs, employing a combined loss function that includes structural similarity (SSIM) and a binary cross-entropy loss to ensure image realism.

The training process involved resampling and co-registering images to a 1x1 mm resolution. Data augmentation techniques like spatial transformations and intensity variations were applied. The GAN was trained on a dataset of 73 patients from two in-house scanners. Crucially, the trained GAN was then used to generate synthetic T2-w fs images for a separate, unseen test dataset of 101 patients from multiple institutions and scanners, simulating a real-world, heterogeneous clinical scenario.

The diagnostic value of these synthetic images was assessed by two neuroradiologists who first graded six different spinal pathologies using only T1-w and non-fs T2-w images. Subsequently, they re-graded the same pathologies with the addition of the synthetic T2-w fs images. The performance was evaluated by comparing these gradings against a "ground truth" (GT) established by consensus of both neuroradiologists, incorporating real T2-w fs images, follow-up scans, and other imaging modalities. Statistical analysis, including Cohen's Kappa (κ) for agreement and accuracy metrics, was used to quantify the improvement.

Datasets

  • Training Data: 73 patients (129 datasets) from two in-house 3T MRI scanners (Philips Healthcare). These datasets included sagittal T1-w TSE, non-fs T2-w TSE, and T2-w fs spine images.
  • Testing Data: 101 patients (101 datasets) from 38 different MRI scanners across three vendors (Philips, Siemens, GE Healthcare) and varying field strengths (1.5T and 3T). These datasets comprised sagittal T1-w TSE and non-fs T2-w TSE scans. The synthetic T2-w fs images were generated from these inputs.

The study utilized retrospective data, and informed consent was waived due to the retrospective nature of the study.

Key Insights and Ideas

  • Bridging the Diagnostic Gap: The most significant insight is that GAN-generated synthetic T2-w fs images can effectively compensate for the absence of actual T2-w fs sequences in routine spinal MRI, thereby improving diagnostic accuracy without increasing scan time.
  • Clinical Feasibility: The generation of synthetic images is remarkably fast, taking less than 5 minutes per dataset, making it a practical addition to the radiological workflow. This speed allows for on-demand generation, potentially integrated into PACS systems.
  • Generalizability: The GAN demonstrated good generalizability, being trained on data from two scanners and successfully generating synthetic images for a diverse test set from 38 different scanners, highlighting its robustness across various acquisition parameters.
  • Enhanced Pathology Detection: The synthetic T2-w fs images specifically aided in the detection of subtle pathologies, such as inflammatory Modic changes and spondylodiscitis-associated fluid collections, which might be missed on T1-w and non-fs T2-w images alone.
  • Simulating Workflow: The study's design, where radiologists graded images first without and then with synthetic T2-w fs, effectively simulated a realistic clinical scenario, providing a more accurate assessment of the synthetic images' utility.

Key Results

  • Improved Agreement with Ground Truth: The addition of synthetic T2-w fs images led to a significantly higher agreement with the ground truth grading compared to using T1-w and non-fs T2-w images alone. The mean Cohen's Kappa (κ) for the synthetic protocol was 0.65, compared to 0.56 for the T1/T2-only protocol (p = 0.043).
  • Higher Accuracy: Accuracy in grading pathologies was generally higher when using the synthetic protocol. For instance, inflammatory Modic changes showed a significant accuracy increase from 80.2% (T1/T2 only) to 87.1% with the synthetic protocol (p = 0.034).
  • Speed: The virtual generation of synthetic T2-w fs images takes less than 5 minutes, with the synthesis itself being under 30 seconds.

Limitations and Future Work

  • Limited Pathology Scope: The study focused on six specific pathologies for which T2-w fs contrast is known to be important. The generalizability of these findings to other spinal pathologies needs further investigation.
  • "Second Look" Effect: The authors acknowledge that the improved grading might be partly due to the radiologists having a "second look" at the images after the synthetic T2-w fs were added. However, they emphasize that the study's intention was to simulate a workflow where radiologists can choose to generate these images.
  • Interrater Variability: Some interrater variability was observed, which the authors attribute to the slight differences in clinical experience between the two neuroradiologists.
  • Future Work: The authors suggest that the demonstrated utility of GANs for data augmentation could be valuable for research settings, especially for analyzing large datasets that may lack T2-w fs sequences. They also highlight the potential for GANs to improve the generalizability of AI models in medical imaging by increasing data diversity. Further research could explore the application of this approach to a broader range of pathologies and investigate its impact on clinical decision-making and patient outcomes.

Summary updated

Abstract

(1) Background and Purpose: In magnetic resonance imaging (MRI) of the spine, T2-weighted (T2-w) fat-saturated (fs) images improve the diagnostic assessment of pathologies. However, in the daily clinical setting, additional T2-w fs images are frequently missing due to time constraints or motion artifacts. Generative adversarial networks (GANs) can generate synthetic T2-w fs images in a clinically feasible time. Therefore, by simulating the radiological workflow with a heterogenous dataset, this study's purpose was to evaluate the diagnostic value of additional synthetic, GAN-based T2-w fs images in the clinical routine. (2) Methods: 174 patients with MRI of the spine were retrospectively identified. A GAN was trained to synthesize T2-w fs images from T1-w, and non-fs T2-w images of 73 patients scanned in our institution. Subsequently, the GAN was used to create synthetic T2-w fs images for the previously unseen 101 patients from multiple institutions. In this test dataset, the additional diagnostic value of synthetic T2-w fs images was assessed in six pathologies by two neuroradiologists. Pathologies were first graded on T1-w and non-fs T2-w images only, then synthetic T2-w fs images were added, and pathologies were graded again. Evaluation of the additional diagnostic value of the synthetic protocol was performed by calculation of Cohen's ĸ and accuracy in comparison to a ground truth (GT) grading based on real T2-w fs images, pre- or follow-up scans, other imaging modalities, and clinical information. (3) Results: The addition of the synthetic T2-w fs to the imaging protocol led to a more precise grading of abnormalities than when grading was based on T1-w and non-fs T2-w images only (mean ĸ GT versus synthetic protocol = 0.65; mean ĸ GT versus T1/T2 = 0.56; <i>p</i> = 0.043). (4) Conclusions: The implementation of synthetic T2-w fs images in the radiological workflow significantly improves the overall assessment of spine pathologies. Thereby, high-quality, synthetic T2-w fs images can be virtually generated by a GAN from heterogeneous, multicenter T1-w and non-fs T2-w contrasts in a clinically feasible time, which underlines the reproducibility and generalizability of our approach.
BibTeX
@inproceedings{schlaeger2023implementation,
  author = {Sarah Schlaeger and Katharina Drummer and Malek El Husseini and Florian Kofler and Nico Sollmann and Severin Schramm and Claus Zimmer and Jan S. Kirschke and Benedikt Wiestler},
  title = {{Implementation of GAN-Based, Synthetic T2-Weighted Fat Saturated Images in the Routine Radiological Workflow Improves Spinal Pathology Detection}},
  booktitle = {Diagnostics},
  year = {2023},
  doi = {10.3390/diagnostics13050974},
}
Figure