Fine-Tuning SAM2 for Coronary Artery Segmentation in X-Ray Fluoroscopy

Technical Summary: Fine-Tuning SAM2 for Coronary Artery Segmentation in X-Ray Fluoroscopy

Problem Statement
While foundation models like SAM2 offer strong general segmentation capabilities, their direct application to medical imaging—specifically X-ray fluoroscopy for coronary artery segmentation—is hindered by domain-specific challenges. These include noise from patient movement, the projection-based nature of X-rays, and low contrast between vessels and the background. Furthermore, existing computer vision models typically require extensive one-to-one training for specific applications, and there is a scarcity of large, publicly available, frame-level annotated datasets for interventional cardiology procedures. The zero-shot application of MedSAM2 to this domain yields negligible performance (Dice score of 0.033), indicating a significant domain gap.

Methodology
The authors propose a fine-tuning strategy for MedSAM2 (a medical-adapted version of SAM2) to address these challenges using a static angiogram dataset for training and a video dataset for qualitative evaluation.

• Model Architecture: The approach utilizes MedSAM2, which employs a Hiera-Tiny backbone, an FPN neck, and a two-way transformer mask decoder. For video processing, it leverages a 6-frame FIFO memory module to propagate masks without re-prompting.
• Datasets:
  • Training: The ARCADE dataset (1,000 training images, 200 validation images) containing polygon annotations for coronary arteries.
  • Evaluation: The CoronaryDominance dataset (10 video studies) used for qualitative video tracking assessment, as it lacks per-frame ground truth masks.
• Fine-Tuning Strategy:
  • Partial Encoder Unfreezing: Instead of training the entire network, the authors unfreeze only the last two Hiera trunk blocks and the FPN neck (approx. 10% of encoder parameters), alongside the mask decoder. This preserves pre-trained medical features while adapting high-level representations.
  • Data Augmentation: Geometric transformations (flips, rotations) are applied to preserve mask integrity, avoiding intensity-based augmentations that could misalign boundaries.
  • Prompting: Training uses a single centroid click derived from ground truth with jitter; inference auto-computes the centroid from the brightest pixels of the prompt frame.
• Loss Function: A composite loss is employed to handle the specific topology of vessels:
  • Dice Loss ($L_{Dice}$) and Weighted BCE ($L_{wBCE}$) to handle class imbalance.
  • clDice Loss ($L_{clDice}$): A topology-preserving loss that operates on the soft skeleton of masks. This penalizes breaks in centerline connectivity, addressing the issue where standard overlap metrics might accept anatomically incorrect, fragmented vessel segments. The weight for clDice is linearly ramped up over the first 8 epochs to stabilize training.
• Video Inference: The model selects the frame with the highest pixel variance (peak contrast) as the prompt frame and propagates the mask forward and backward using the streaming memory architecture.

Key Results

• Static Segmentation (ARCADE Validation):
  • Zero-Shot: MedSAM2 achieved a Dice score of 0.033, confirming the domain gap.
  • Frozen Encoder Fine-Tuning: Unfreezing only the decoder improved the score to 0.727.
  • Full Fine-Tuning (Ours): Partially unfreezing the encoder, adding clDice loss, and using augmented data achieved a Dice score of 0.767 ± 0.082 (IoU 0.629). This performance is competitive with models specifically designed for X-ray coronary angiography (XCA).
• Video Inference (CoronaryDominance):
  • The fine-tuned model tracks the main coronary trunk coherently across frames without re-prompting.
  • Specificity: In 9 out of 10 video studies, the fine-tuned model successfully avoided false positives on ribs, surgical stents, bypass grafts, and implanted devices. In contrast, the zero-shot model consistently mislabeled these high-contrast, tubular structures as vessels.
  • Limitations: Both models fail when contrast dye clears between injection cycles, as the memory module loses the target. Re-prompting at the start of each injection cycle is required.

Significance and Claims
The paper claims that MedSAM2 can effectively adapt to X-ray fluoroscopy with minimal data (1,000 images) and without extensive retraining of the entire backbone. The primary significance lies in the specificity of the fine-tuned model: it learns the genuine appearance of coronary vessels in this modality rather than just detecting generic tubular shapes. This reduces clinically dangerous false positives (e.g., identifying stents as vessels) without requiring explicit negative supervision.

The authors note that while the static segmentation results are quantitative, the video evaluation remains qualitative due to the lack of frame-level ground truth in available public datasets. They acknowledge that the current memory-based architecture introduces tracking latency at the start of contrast injection cycles and that future work should explore self-supervised pre-training on unlabeled video and deployment on resource-constrained clinical hardware.