Using Foundation Models as Pseudo-Label Generators for Pre-Clinical 4D Cardiac CT Segmentation

TL;DR

This work tackles the domain gap between human and preclinical porcine cardiac CT by using foundation models trained on human data to generate initial pseudo-labels for porcine 4D CT. It then applies a simple, iterative self-training strategy, where a refined model re-labels frames and feeds updated pseudo-labels back into training, aiming to improve both segmentation accuracy and temporal consistency. Across experiments, the approach yields meaningful improvements over initial pseudo-labels, with high performance in human CT validation and notable gains in temporal stability for porcine data, though artifacts and cross-species differences limit complete parity with manual labels. The study demonstrates a label-efficient pathway for preclinical cardiac imaging analysis, supporting translational research while highlighting avenues for more sophisticated self-training and multi-model fusion in future work.

Abstract

Cardiac image segmentation is an important step in many cardiac image analysis and modeling tasks such as motion tracking or simulations of cardiac mechanics. While deep learning has greatly advanced segmentation in clinical settings, there is limited work on pre-clinical imaging, notably in porcine models, which are often used due to their anatomical and physiological similarity to humans. However, differences between species create a domain shift that complicates direct model transfer from human to pig data. Recently, foundation models trained on large human datasets have shown promise for robust medical image segmentation; yet their applicability to porcine data remains largely unexplored. In this work, we investigate whether foundation models can generate sufficiently accurate pseudo-labels for pig cardiac CT and propose a simple self-training approach to iteratively refine these labels. Our method requires no manually annotated pig data, relying instead on iterative updates to improve segmentation quality. We demonstrate that this self-training process not only enhances segmentation accuracy but also smooths out temporal inconsistencies across consecutive frames. Although our results are encouraging, there remains room for improvement, for example by incorporating more sophisticated self-training strategies and by exploring additional foundation models and other cardiac imaging technologies.

Figures (4)

• Figure 1: Self-training for 4D CT data, using a foundation model to initialize pseudo labels, with an (optional) additional small labeled dataset.
• Figure 2: Fraction of flagged segmentations for each iteration, with iteration 1 = the initial foundation model predictions. A: The models were trained on TotalSegmentator pseudo labels only, B: The models were trained on a mix of TotalSegmentator pseudo labels and manually labeled human data. C: The models were trained in a standard self-training fashion where the model itself provides the initial pseudo labels.
• Figure 3: Segmentation results using the "pseudo only" model on porcine data, comparing the initial pseudo-labels (middle) and final pseudo-labels (right). Label abbreviations: RA: right atrium, LA: left atrium, RV: right ventricle, LV: left ventricle, LV myo: left ventricle myocardium, AO: aorta, PA: pulmonary artery. The plots on the right show the temporal consistency of the left ventricle volume across frames, different colors represent different imaging visits of the same subject. A: An example with an already accurate initial segmentation, resulting in minimal changes after self-training (frame 0 of the pink curve). B: A case showing improvements through self-training (frame 4 of the pink curve). C: Another case with large changes following self-training (frame 7 of the blue curve). D: An example where slight improvements are achieved but some errors persist (frame 4 of the pink curve). E: A case with an LV catheter that is causing artifacts. (frame 5 of the purple curve).
• Figure 4: Fraction of flagged segmentations and example segmentation of TotalSegmentator and our "pseudo only" model of unseen data.