Source-linked AI summary
LiFUSE: Frequency-Spatial Unified Self-Supervised Pretraining for Liver Disease Analysis on Non-contrast CT
Hao Zhang, Yi Huang, Wei Liu, Yuan Gao, Zeju Li, Xu Han, Yu Shi, Ling Zhang, Ke Yan
TL;DR
The growing burden of liver tumors and diffuse liver disease makes effective screening and diagnosis important, and NCCT is a lower-cost, safer option for opportunistic screening. LiFUSE combines frequency- and spatial-domain pretraining with cross-domain consistency. It consistently outperformed strong supervised and self-supervised baselines across six downstream tasks, including transfer to a public target-domain cohort.
Problem
The growing burden of liver tumors and diffuse liver disease makes effective screening and diagnosis important, while NCCT is used as a safer, lower-cost option for opportunistic screening.
Source: §1 · pp. 1–2Method
Results
LiFUSE consistently outperformed strong supervised and self-supervised baselines across six downstream tasks evaluated on internal cohorts and a public target-domain cohort.
Source: §4 · p. 9Takeaways & Limitations
The results suggest potential for large-scale opportunistic liver screening using NCCT.
Source: §4 · p. 9
Abstract
from arXiv · showhide
Given the growing global burden of liver tumors and diffuse liver diseases, non-contrast computed tomography (NCCT) is increasingly used for opportunistic screening because of its low cost and favorable safety profile. However, downstream tasks such as tumor detection and fine-grained disease diagnosis remain challenging due to NCCT's inherently low tissue contrast, lack of hemodynamic enhancement, and complex anatomical background. These limitations make subtle, heterogeneous, and multifocal hepatic lesions difficult to identify. Although large-scale self-supervised pretraining can learn robust representations from unlabeled NCCT scans, conventional voxel-wise reconstruction does not fully exploit the complementary information available in NCCT. Under weak NCCT contrast, low-level intensity patterns become less informative for pathology recognition, limiting the capture of lesion-specific discriminative features. In this paper, we propose the Liver Frequency-Spatial Unified Self-Supervised Pretraining (LiFUSE). The frequency branch applies random band-stop masking and spectrum reconstruction to encourage the model to capture subtle differences between lesions and surrounding tissues, while the spatial branch performs hybrid-granularity masked volumetric reconstruction to accommodate lesions of diverse sizes and morphologies. Finally, a cross-domain consistency scheme couples the spatial and frequency branches, yielding more discriminative representations for hepatic structures and lesions. Extensive experiments on three clinically relevant tasks across internal cohorts and a public target-domain transfer setting show that LiFUSE consistently outperforms strong baselines. These results suggest that unified frequency-spatial learning can improve liver disease analysis on NCCT, highlighting its potential for large-scale screening.
1 Introduction
Rising liver-disease burden motivates screening, but NCCT is challenging for subtle lesions, and spatial or frequency information alone may not distinguish them. LiFUSE couples both domains and is reported to outperform state-of-the-art supervised and self-supervised methods across its evaluated clinical settings.
1 Introduction
- NCCT offers a safer, lower-cost option for opportunistic screening, while subtle, heterogeneous, and multifocal liver lesions remain difficult to identify under low soft-tissue contrast.The paper frames screening and diagnosis as important amid rising liver cancer and diffuse liver disease burden.Source: §1 · pp. 1–2
- Voxel-wise reconstruction focuses on spatial intensity and texture, while spatial-only or frequency-only statistics overlap between normal tissue and lesions.The paper argues that frequency information may add complementary structural and pathological cues, but explicit cross-domain coupling is needed for joint optimization.Source: §1 · p. 2, Figure 1 · p. 2
- LiFUSE consistently outperformed state-of-the-art supervised and self-supervised methods across internal cohorts and a public target-domain transfer setting.The evaluations covered liver tumor screening, subtype classification, and diffuse liver disease diagnosis.Source: §1 · pp. 2–3
- LiFUSE combines hybrid-granularity masking across frequency and spatial domains with parallel dual-path reconstruction and cross-domain consistency learning.The masking is designed to accommodate diverse lesion morphologies while preserving cross-domain consistency.Source: §1 · p. 3, Figure 2 · p. 3
2 Methodology
LiFUSE pretraining combines frequency-spatial hybrid-granularity masking, parallel dual-path reconstruction, and frequency-spatial unified learning.
2.1 Frequency-Spatial Hybrid-Granularity Mask
- Band-stop masking suppresses a contiguous spectral band, then spatial masking applies the same ratio to scale-2 cubes only within scale-1 masked regions.The scale-1 crop is partitioned into 2 × 2 × 2 cubes to accommodate lesion sizes and morphologies.Source: §2.1 · pp. 3–4
2.2 Parallel Dual-Path Reconstruction
- A shared encoder feeds spatial voxel reconstruction and complex-spectrum reconstruction in parallel, preserving spatial appearance cues alongside frequency magnitude and phase.The spatial branch uses masked-region MSE and a cross-granularity consistency loss aligning two-scale reconstructions over their shared masked region.Source: §2.2 · p. 4
- The frequency branch applies focal frequency loss to adaptively reweight spectral components, emphasizing those harder to fit.The complex spectrum preserves both magnitude and phase information.Source: §2.2 · pp. 4–5
2.3 Frequency-Spatial Unified Learning
- Unified learning FFT-transforms spatial predictions and inverse-transforms frequency predictions, then enforces consistency with the original spectrum and volume using a β-weighted objective.The frequency-spatial unified learning weight β is a trade-off coefficient.Source: §2.3 · p. 5
3 Experiments
Across six downstream tasks, LiFUSE is evaluated on internal NCCT cohorts and a public target-domain cohort, with comparisons, ablations, masking-order tests, and hyperparameter analyses. It outperforms competing methods on internal tumor tasks and public-cohort transfer after target-domain fine-tuning.
3.1 Datasets
- The cohorts include 7,598 unlabeled pretraining scans and labeled sets for tumor screening and subtype classification, internal diffuse disease diagnosis, and public steatosis transfer.Cases were separated patient-wise across pretraining, fine-tuning, validation, and test sets.Source: §3.1 · pp. 5–6
3.3 Experimental Results
- LiFUSE achieves the best internal-test performance, improving AUC by 0.51% for tumor screening and 1.28% for subtype classification over the best competing method.The comparisons cover liver tumor screening and fine-grained subtype classification.Source: Table 1 · p. 6, §3.3 · p. 7
- On the public UNIFESP target-domain test set, LiFUSE improves AUC by 1.39% over the best baseline after target-domain fine-tuning.This evaluates fine-tuned transfer, not adaptation-free external generalization.Source: Table 2 · p. 6, §3.3 · p. 7
3.4 Ablation Study
- In ablations, adding cross-domain coupling improves tumor-screening AUC by 1.25% over the variant without unified learning, while joint spatial-frequency modeling beats either single branch.LiFUSE gains 1.15% and 0.38% AUC over spatial-only and frequency-only variants, respectively.Source: Table 3 · p. 7, §3.4 · p. 8
3.2 Implementation Details
- LiFUSE’s unified feature space shows more distinct clusters and lower intra-class variance than single-domain embeddings; evaluation uses AUC, sensitivity, and specificity.Binary-task thresholds are selected on validation data using the Youden index, while test sets are reserved for final evaluation.Source: Figure 3 · p. 7, §3.2 · p. 7
3.5 Analysis of Masking Order
- Frequency-then-spatial masking performs better than the reverse on screening, with higher sensitivity at comparable specificity, and raises subtype average AUC from 95.48% to 96.28%.The proposed order preserves spectral structure before spatial patching introduces localized perturbations.Source: Table 4 · p. 8, §3.5 · p. 8
3.6 Analysis of Hyper-parameters
- The tested defaults are α=0.2, β=0.3, mask sizes (4, 8, 16), and masking ratio µ=75%; the medium mask sizes outperform smaller and larger settings in tumor screening.β=0.3 improves malignant screening by 2.76% over β=0.1, while µ=75% performs best in malignant screening among the tested ratios.Source: Figure 4 · p. 8, §3.6 · pp. 8–9
4 Discussion and Conclusion
LiFUSE couples frequency-domain spectral masking and spatial-domain patch masking under cross-domain consistency constraints. Evaluations on internal cohorts and supervised transfer to a public target-domain cohort show consistent gains over strong baselines across six downstream tasks, supporting its potential for opportunistic liver screening.
4 Discussion and Conclusion
- LiFUSE combines frequency and spatial masking with cross-domain consistency, and outperforms strong supervised and self-supervised baselines across six downstream tasks.The authors identify multimodal auxiliary information and adaptation to other solid organs as future directions.Source: §4 · p. 9