Abstract
Accurate liver tumor diagnosis typically relies on Contrast-Enhanced MRI (CEMRI), which requires the injection of contrast agents that can be toxic and expensive. We present EaMtNet (Edge-aware Multi-task Network), a novel framework for the simultaneous segmentation, clinical index quantification, and uncertainty estimation of liver tumors using only non-contrast MRI (NCMRI). By leveraging edge-aware Transformers and multi-task learning, EaMtNet overcomes the boundary blurriness inherent in NCMRI. Our results show a Dice Similarity Coefficient of 90.01% and a Mean Absolute Error of 2.72 mm for maximum diameter, significantly outperforming current state-of-the-art models.
Introduction
The clinical standard for liver tumor assessment often involves CEMRI to achieve clear tumor visualization. However, contrast agents pose risks to patients with renal impairment and increase healthcare costs. Non-contrast MRI (NCMRI) is a safer alternative but is difficult to interpret due to poor boundary definition. Furthermore, existing AI models typically address segmentation and quantification as separate tasks, missing opportunities for synergistic learning.
EaMtNet addresses these challenges by integrating edge-aware feature aggregation and multi-task prediction heads into a unified architecture. It aims to provide an "all-in-one" tool that outlines the tumor, measures its size, and estimates the reliability of its own predictions.
Methods
The EaMtNet architecture follows a three-step process:
- Dual-Stream Feature Extraction: The model uses two parallel CNN encoders (based on a UNet backbone) to process T2FS and DWI images separately. Simultaneously, Sobel filters are applied to extract edge maps, providing the network with explicit "prior knowledge" about potential tumor boundaries.
- Edge-aware Feature Aggregation (EaFA): This module uses a Transformer-based mechanism to fuse features from both MRI modalities and the edge maps. The self-attention mechanism allows the network to capture long-range dependencies, helping the model understand how a blurry region in one image relates to a sharp edge detected in another.
- Multi-task Prediction: The fused features are passed to three specialized heads:
- Quantification Head: A linear layer for image-level regression (Max Diameter, Center Points, Area).
- Segmentation Head: A decoder that outputs a pixel-level mask.
- Uncertainty Head: Uses an entropy-based measure to generate a map indicating where the network is most "confused," typically at the tumor boundaries.
The model is trained using a cross-task joint loss function, allowing the segmentation and quantification tasks to mutually improve one another.
Results and Discussion
The study utilized a clinical dataset of 250 subjects, with ground truth verified by senior abdominal radiologists. EaMtNet significantly outperformed state-of-the-art (SOTA) models like TransUNet and UNet++:
- Segmentation: Achieved a Dice Similarity Coefficient (DSC) of 90.01%, a 5.4% improvement over the best SOTA baseline.
- Quantification: The Mean Absolute Error (MAE) for the Maximum Diameter (MD) was reduced to just 2.72 mm.
- Uncertainty Estimation: By outputting an entropy map, the system provides a visual "safety check" for clinicians. High entropy at a tumor's edge signals that the AI's measurement might require manual verification.
Ablation studies confirmed that the EaFA module is critical; its removal caused the DSC to drop from 90.01% to 85.82%, proving that the edge-aware Transformer fusion is essential for interpreting non-contrast images.
Conclusion
EaMtNet provides a robust, contrast-free solution for liver tumor analysis. By effectively integrating edge information and multi-task learning, it matches or exceeds the diagnostic utility of contrast-dependent methods while enhancing patient safety. While currently limited to 2D axial slices, future work will explore 3D volumetric extensions and multi-center validation to ensure broader clinical applicability.
Summary updated