Overview
Liver cancer ranks as the sixth most prevalent cancer worldwide and is a significant contributor to cancer-related mortality. Accurate segmentation of liver tumors is essential for effective disease management. However, traditional manual segmentation by radiologists is often time-consuming and varies based on the radiologist’s expertise.
Advancements in AI Technology
Recent developments in artificial intelligence (AI) have led to the creation of tumor segmentation models that utilize deep convolutional neural networks. These models can identify and outline the precise shape, size, and location of tumors in medical imaging scans. However, they typically require large datasets, often between 1,000 to 10,000 cases, which poses a challenge in medical AI.
Introducing MHP-Net
A research team from the Institute of Science Tokyo, led by Professor Kenji Suzuki and PhD student Yuqiao Yang, has developed a novel AI model named MHP-Net. This model can accurately segment liver tumors from computed tomography (CT) scans, even when trained on very small datasets, outperforming existing state-of-the-art systems. Their findings were published in the journal IEEE Access on May 16, 2025.
How MHP-Net Works
- MHP-Net employs a multi-scale Hessian-enhanced patch-based neural network architecture.
- The model divides medical images into small 3D patches, allowing it to focus on individual sections rather than the entire image.
- Each patch is paired with an enhanced version created through Hessian filtering, which emphasizes spherical objects like tumors.
Performance Evaluation
The effectiveness of MHP-Net was assessed using the Dice similarity score, which measures the accuracy of predicted segmentations against expert-annotated images. The model achieved impressive scores of:
- 0.691 with 7 tumors
- 0.709 with 14 tumors
- 0.719 with 28 tumors
These results indicate that MHP-Net surpasses established models such as U-Net, Res U-Net, and HDense-U-Net.
Advantages of MHP-Net
- Fast training time of under 10 minutes.
- Real-time inference capability of approximately 4 seconds per patient.
- Lightweight architecture suitable for clinical settings with limited computational resources.
Future Implications
Professor Suzuki emphasizes that this research marks the beginning of small-data AI in medical imaging, where effective deep learning models can be developed from limited datasets. MHP-Net’s success could pave the way for similar solutions in other medical imaging areas, including the detection of rare cancers.
This study highlights the potential of small-data AI in medical image analysis, democratizing access to advanced AI technologies in under-resourced healthcare settings. The researchers aim to explore broader applications of small-data AI models, facilitating scalable and cost-effective AI deployment in healthcare globally.
For further details, refer to the study: Patch-Based Deep-Learning Model With Limited Training Dataset for Liver Tumor Segmentation in Contrast-Enhanced Hepatic Computed Tomography, published in IEEE Access.