AI Model Achieves Accurate Liver Tumor Segmentation With Minimal Data

MHP-Net: Accurate Liver Tumor Segmentation with Limited Data

Liver cancer ranks as the sixth most common cancer worldwide and remains a leading cause of cancer-related deaths. Effective management relies heavily on accurately identifying and segmenting liver tumors in medical images. However, manual segmentation by radiologists is time-consuming and prone to variability depending on expertise.

Artificial intelligence (AI) models, particularly those based on deep convolutional neural networks, have improved tumor segmentation by automatically outlining tumor boundaries in medical scans. Yet, these models typically require large datasets—often thousands of cases—to perform well. This data dependency poses a significant challenge in medical AI, where acquiring extensive labeled datasets is difficult.

Introducing MHP-Net

A research team from the Biomedical AI Research Unit at the Institute of Science Tokyo, led by Professor Kenji Suzuki and PhD student Yuqiao Yang, developed an AI model called the multi-scale Hessian-enhanced patch-based neural network (MHP-Net). This model achieves high segmentation accuracy using very small training datasets, outperforming existing state-of-the-art methods.

MHP-Net processes 3D computed tomography (CT) images by dividing them into small patches. Each patch is paired with a Hessian-filtered version, which enhances spherical structures like tumors. This dual-input approach helps the model focus on relevant features, resulting in precise tumor segmentation maps from contrast-enhanced CT scans.

Performance and Efficiency

The team evaluated MHP-Net using the Dice similarity score, a metric measuring overlap between predicted and expert-annotated tumor regions (ranging from 0 to 1). Despite training on datasets as small as 7, 14, and 28 tumors, the model achieved Dice scores of 0.691, 0.709, and 0.719 respectively. These results exceed the performance of popular models like U-Net, Res U-Net, and HDense-U-Net.

Besides accuracy, MHP-Net offers fast training times (under 10 minutes) and real-time inference speed (~4 seconds per patient). Its lightweight design makes it practical for clinical environments with limited computational resources.

Implications for Medical AI

MHP-Net demonstrates the potential of small-data AI approaches in medical imaging, reducing the need for extensive labeled datasets. This can democratize AI adoption in healthcare settings, especially in under-resourced hospitals and clinics.

The researchers plan to expand the application of small-data AI models to other medical imaging challenges, including the detection of rare cancers. This approach could lead to scalable, cost-effective AI tools that assist clinicians worldwide.

About Institute of Science Tokyo (Science Tokyo)

Established on October 1, 2024, the Institute of Science Tokyo was formed through the merger of Tokyo Medical and Dental University and Tokyo Institute of Technology. Its mission is to advance science and human wellbeing to create value for society.

• Journal: IEEE Access
• DOI: 10.1109/ACCESS.2025.3570728
• Method of Research: Experimental study
• Article Title: Patch-Based Deep-Learning Model With Limited Training Dataset for Liver Tumor Segmentation in Contrast-Enhanced Hepatic Computed Tomography
• Publication Date: 16-May-2025