Efficiency vs. Efficacy: Assessing the Compression Ratio-Dice Score Relationship through a Simple Benchmarking Framework for Cerebrovascular 3D Segmentation

The increasing size and complexity of medical imaging datasets, particularly in 3D formats, present significant barriers to collaborative research and transferability. This study investigates whether the ZFP compression technique can mitigate these challenges without compromising the performance of automated cerebrovascular segmentation, a critical first step in intracranial aneurysm detection. We apply ZFP in both its error tolerance and fixed-rate modes to a large scale, and one of the most recent, datasets in the literature, 3D medical dataset containing ground-truth vascular segmentations. The segmentation quality on the compressed volumes is rigorously compared to the uncompressed baseline (Dice approximately equals 0.8774). Our findings reveal that ZFP can achieve substantial data reduction–up to a 22.89:1 ratio in error tolerance mode–while maintaining a high degree of fidelity, with the mean Dice coefficient remaining high at 0.87656. These results demonstrate that ZFP is a viable and powerful tool for enabling more efficient and accessible research on large-scale medical datasets, fostering broader collaboration across the community.

💡 Research Summary

This paper investigates whether the ZFP (Zero‑Error Fixed‑Precision) lossy compression algorithm can substantially reduce the storage and transmission burden of large‑scale 3D cerebrovascular imaging data without degrading the performance of downstream automated segmentation, a key step in intracranial aneurysm detection. The authors use the RSNA Intracranial Aneurysm Detection dataset, which comprises multi‑site CT and MR volumes together with expert‑annotated masks of 13 vascular territories. The uncompressed dataset occupies roughly 3.85 GB.

Two ZFP operating modes are examined: (1) fixed‑rate, where a constant number of bits per voxel is allocated (16, 8, 4, 2 bits/voxel), and (2) error‑tolerance, where the decompressed data are guaranteed to stay within an absolute error bound (500, 1000, 1500). Compression is performed with the Python wrapper pyzfp on a single NVIDIA A100 GPU.

For segmentation, the authors adopt two state‑of‑the‑art Mamba‑based networks—MambaVesselNet and U‑Mamba—both of which combine convolutional feature extraction with long‑range modeling via state‑space sequence models. Models are trained on full‑resolution volumes using a 5 000‑iteration schedule, 64³ voxel patches, batch size 8, Adam optimizer (initial LR 1e‑4, cosine annealing to 1e‑7), and an adaptive loss that transitions from Dice + Cross‑Entropy to Focal Loss. Performance is quantified with Dice Similarity Coefficient (DSC) and Intersection‑over‑Union (IoU).

Results show an almost linear trade‑off between compression ratio and segmentation fidelity. In fixed‑rate mode, reducing the bit depth from 16 bits/voxel (compression ratio ≈ 1:1) to 2 bits/voxel (≈ 8:1) only lowers the mean Dice from 0.87738 to 0.87734—a negligible 0.00004 drop. In error‑tolerance mode, an absolute error bound of 1500 yields a compression ratio of 22.89:1 while the mean Dice remains high at 0.87656, a reduction of merely 0.001 relative to the uncompressed baseline. IoU follows a similar pattern, confirming that even aggressive lossy compression does not materially affect the delineation of thin vascular structures. Visual quality metrics (PSNR, SSIM) degrade substantially, underscoring that task‑oriented quality (segmentation accuracy) can be preserved despite poor perceptual fidelity.

The authors argue that ZFP offers several practical advantages over learned compressors (e.g., MedZip, TDSIC): no training data or time is required, implementation is straightforward, and compression/decompression can be performed on CPU or modest GPU resources. Their experiments demonstrate that ZFP can achieve compression ratios up to 50:1 while keeping Dice > 0.86, establishing it as a viable, cost‑effective alternative for large‑scale medical imaging pipelines.

Limitations include a relatively narrow exploration of compression parameters and the absence of direct comparisons with other modern codecs (e.g., JP3D, wavelet‑based, or implicit neural representations). The study also focuses on a single segmentation architecture, leaving open the question of how universally robust the findings are across different network designs or multi‑modal inputs.

Future work is suggested in three directions: (1) extending the analysis to additional segmentation models and to multi‑modal (CT + MRI) data to assess generalizability; (2) integrating compression parameter selection into an automated meta‑learning or reinforcement‑learning framework to tailor the optimal trade‑off for each dataset; and (3) evaluating end‑to‑end clinical workflows, including real‑time compression/decompression latency and secure data sharing mechanisms.

In conclusion, the study provides strong empirical evidence that ZFP compression can dramatically shrink 3D cerebrovascular datasets while preserving the accuracy of downstream deep‑learning segmentation. This finding has immediate implications for collaborative research, cloud‑based model training, and the broader goal of democratizing access to high‑resolution medical imaging data.