Abstract
Hepatocellular carcinoma (HCC), ranking as the third leading cause of cancer-related mortality worldwide, demands urgent improvements in early detection to enhance patient survival. While ultrasound remains the preferred screening modality due to its cost-effectiveness and real-time capabilities, its sensitivity (59%-78%) heavily relies on radiologists' expertise, leading to inconsistent diagnostic outcomes and operational inefficiencies. Recent advancements in AI technology offer promising solutions to bridge this gap. This study introduces the Hierarchical Sparse Query Transformer (HSQformer), a novel hybrid architecture that synergizes CNNs' local feature extraction with Vision Transformers' global contextual awareness through latent space representation and sparse learning. By dynamically activating task-specific experts via a Mixture-of-Experts (MoE) framework, HSQformer achieves hierarchical feature integration without structural redundancy. Evaluated across three clinical scenarios: single-center, multi-center, and high-risk patient cohorts, HSQformer outperforms state-of-the-art models (e.g., 95.38% AUC in multi-center testing) and matches senior radiologists' diagnostic accuracy while significantly surpassing junior counterparts. These results highlight the potential of AI-assisted tools to standardize HCC screening, reduce dependency on human expertise, and improve early diagnosis rates.
Introduction
Hepatocellular carcinoma (HCC) is a leading cause of cancer death globally. While B-mode ultrasound is the primary screening tool due to its cost-effectiveness and real-time nature, its effectiveness is hampered by a "sensitivity gap." Diagnostic accuracy currently ranges from 59% to 78% because it relies heavily on a radiologist’s subjective expertise. This leads to inconsistent outcomes and missed early-stage cases.
Existing AI solutions often struggle to balance the need for local texture analysis (where CNNs excel) with the need for global anatomical context (where Vision Transformers excel). Simply "gluing" these two architectures together (late fusion) often yields marginal benefits because it tends to overemphasize high-level semantics while ignoring the fine-grained details crucial for medical diagnosis.
Methods
The Hierarchical Sparse Query Transformer (HSQformer) employs a "best of both worlds" strategy using several novel components to address these challenges:
- Hybrid Feature Extraction: It uses both ConvNext (CNN) and SwinTransformer (ViT) as backbones to capture multi-scale representations simultaneously.
- Latent Space Projection: Instead of processing raw features, a projector module maps these multi-scale features into a compact $D$-dimensional latent space. This reduces redundancy and focuses the model on the most salient diagnostic features.
- Hierarchical Sparse Querying: Unlike standard Q-formers that only query the final high-level semantic layer, HSQformer uses a hierarchical strategy. It injects learnable queries across four stages, capturing both low-level textures and high-level semantics.
- Mixture-of-Experts (MoE): Within its Cross-Self-attention Mixed experts (CSM) blocks, the model dynamically routes tokens to specific "expert" MLP networks. By using a Top-k routing strategy (activating only the most relevant experts), the model increases its capacity and specialization without a massive increase in computational cost during inference.
Datasets and Evaluation
The study utilized the UE-MICAP dataset, a large-scale multi-center collection from the Ultrasound Engineering Society of China, comprising 11,149 cases with 19,464 images collected between 2014 and 2023. Data was sourced from eight different hospitals to ensure diversity. The model was rigorously tested across three distinct clinical scenarios:
- Single-center test: Evaluation on a dataset separate from the training set but from the same institution.
- Multi-center test: Evaluation across diverse clinical environments to test generalizability.
- High-risk patient test: Specifically focusing on patients with hepatitis, a challenging subgroup for HCC screening.
Key Results
HSQformer demonstrated superior performance across all testing scenarios:
- State-of-the-Art Performance: In multi-center testing, HSQformer-B achieved a leading AUC of 95.38% and an accuracy of 88.29%, outperforming standard models like ResNet101, ViT, and ConvNext.
- Human-Machine Parity: In a head-to-head comparison with six radiologists, HSQformer significantly outperformed junior radiologists (3-5 years experience) and matched the diagnostic accuracy of senior radiologists (10+ years experience).
- High Sensitivity for Screening: In the high-risk patient cohort, the model achieved a Recall of 94.09%. This is particularly important for screening tools, where missing a potential cancer case (a false negative) is the most critical error.
Conclusion and Future Work
HSQformer represents a significant step toward standardizing HCC screening. By effectively integrating local and global features through a hierarchical sparse querying mechanism, it achieves clinical-grade performance. Future work will focus on improving model interpretability, optimizing hyperparameter sensitivity, and developing more compact versions of the architecture to facilitate deployment in resource-constrained clinical settings.
Summary updated