Using Foundation Models as Pseudo-label Generators for Pre-clinical 4D Cardiac CT Segmentation

Anne-Marie Rickmann, Stephanie Thorn, Shawn S. Ahn, Supum Lee, Selen Uman, Taras Lysyy, Rachel Burns, Nicole Guerrera, Francis G. Spinale, Jason A. Burdick, Albert J. Sinusas, James S. Duncan

Introduction

In cardiovascular research, porcine models are indispensable due to their anatomical and physiological similarities to humans. These models allow researchers to study disease progression and test interventions in a controlled environment. However, the transition from pre-clinical research to automated analysis is often hindered by a lack of labeled data. While large-scale, annotated datasets exist for human cardiac imaging, equivalent resources for porcine data are scarce. This "data gap" makes training robust deep learning models for animal studies a significant challenge.

Furthermore, cardiac imaging is inherently four-dimensional (4D), consisting of 3D volumes captured over the cardiac cycle. Standard segmentation approaches often process these time frames independently, which frequently results in temporal inconsistencies—visual "jitter" or unrealistic volume fluctuations between heartbeats. This paper presents a novel framework that leverages human foundation models to segment porcine cardiac CT scans without requiring any manual animal-specific annotations.

Methodology: Iterative Self-Training

The core of the proposed method is an iterative self-training pipeline designed to bridge the domain gap between human and porcine anatomy. The process requires zero manual labels for the target species, instead relying on a "warm start" from existing human-trained models.

1. Initial Pseudo-Labeling

The researchers utilize TotalSegmentator, a foundation model trained on thousands of human CT scans. Despite the anatomical differences between species, the model possesses enough general knowledge of cardiac structures to generate initial, albeit noisy, segmentation masks for the porcine data.

2. The Refinement Loop

These initial masks serve as "pseudo-labels" for a second architecture, nnU-Net. The training proceeds in an iterative fashion:

  • Training: The nnU-Net is trained on the pseudo-labels generated in the previous step.
  • Prediction: The newly trained model predicts updated labels for the same porcine dataset.
  • Iteration: This process is repeated for five iterations. With each round, the model effectively "denoises" the labels, adapting to the specific contrast, morphology, and artifacts present in the porcine scans.

3. Enhancing Temporal Consistency

By training on all frames of the cardiac cycle simultaneously, the model learns the underlying temporal logic of the beating heart. Even when processing frames individually during inference, the weights learned during the self-training process encourage smoother transitions and more realistic volumetric changes over time.

Experimental Setup and Datasets

The study utilized a diverse range of data to test the robustness of the pipeline:

  • Porcine Data: 126 3D sequences from 28 pigs (totaling 1,249 frames), including both healthy subjects and those with myocardial infarctions.
  • Human Data (MMWHS): 20 human 3D CT scans from the Multi-Modality Whole Heart Segmentation challenge were used for validation and to establish a mixed-training baseline.
  • Unseen Test Set: A challenging set of 13 pigs (371 frames) featuring thoracotomy (open-chest surgery) artifacts and surgical catheters.

Key Insights and Results

Overcoming Domain Shift

The study demonstrates that human foundation models are sufficiently "species-agnostic" to provide a viable starting point for animal models. While the initial human-trained model produced errors—such as hallucinating human-like branching on a pig’s aorta—the iterative self-training process allowed the model to converge on anatomically plausible porcine structures.

Quantitative Improvements

Because ground-truth labels for the porcine data were unavailable, the authors employed "plausibility" metrics to evaluate performance. These metrics flagged segmentations with disconnected components or unrealistic volume spikes.

  • Error Reduction: The fraction of flagged (implausible) segmentations dropped from approximately 15% using the initial foundation model to between 3% and 6% after five iterations.
  • Temporal Stability: The refined model showed significantly lower frame-to-frame Dice variance compared to the baseline. Volume curves for the heart chambers became smoother, reflecting a more realistic representation of cardiac motion.

Robustness to Artifacts

The method proved resilient even in the presence of surgical artifacts. In the unseen test set involving open-chest procedures, the self-trained model outperformed the direct application of the human foundation model, successfully navigating the presence of catheters and altered anatomical presentations.

Conclusion and Future Directions

This research provides a scalable pathway for analyzing pre-clinical imaging data without the bottleneck of manual annotation. By treating foundation models as pseudo-label generators, researchers can adapt powerful AI tools to specialized domains with minimal human intervention.

Future work may focus on "uncertainty-aware" self-training, where the model selectively learns from the most confident pseudo-labels. Additionally, refining the segmentation of thin, complex structures like the pulmonary artery remains a priority to further improve the clinical utility of these automated porcine heart models.

Summary updated

BibTeX
@inproceedings{rickmann2025using,
  author = {Anne-Marie Rickmann and Stephanie Thorn and Shawn S. Ahn and Supum Lee and Selen Uman and Taras Lysyy and Rachel Burns and Nicole Guerrera and Francis G. Spinale and Jason A. Burdick and Albert J. Sinusas and James S. Duncan},
  title = {{Using Foundation Models as Pseudo-label Generators for Pre-clinical 4D Cardiac CT Segmentation}},
  booktitle = {FIMH},
  year = {2025},
  doi = {10.1007/978-3-031-94562-5\_23},
}
Figure