A Coarse-to-Fine Fusion Network for Small Liver Tumor Detection and Segmentation: A Real-World Study.

Area of Science:

• Medical image analysis
• Radiology
• Artificial intelligence in medicine

Background:

• Accurate liver tumor segmentation is vital for diagnosis and treatment planning.
• Existing methods struggle with segmenting small liver tumors ( < 3.0 cm).
• Multi-modal MRI data is essential for comprehensive liver tumor analysis.

Purpose of the Study:

• To develop a novel coarse-to-fine fusion segmentation approach for improved detection and segmentation of small liver tumors.
• To enhance segmentation accuracy for various liver tumor types, including hepatocellular carcinoma (HCC) and metastases.
• To evaluate the proposed method against conventional and fusion-based segmentation networks.

Main Methods:

• A coarse-to-fine fusion segmentation strategy incorporating a detection module and a CSR (convolution-SE-residual) module.
• Utilizing a private liver MRI dataset with 3605 tumors from 218 patients, focusing on small tumors (< 3.0 cm).
• Comparative analysis against 3D UNet, nnU-Net, and nnDetection integrated fusion networks.

Main Results:

• The proposed method achieved an average Dice similarity coefficient (DSC) of 86.9% and recall of 86.7% on a test set.
• Demonstrated superior performance over single-stage networks and existing fusion approaches.
• Achieved state-of-the-art results for segmenting small objects (< 10 mm) with a DSC of 85.3% and malignancy detection rate of 87.5%.

Conclusions:

• The novel coarse-to-fine fusion segmentation approach effectively detects and segments small liver tumors with high accuracy.
• The method shows significant improvements, particularly for small-sized tumors, offering a promising tool for clinical applications.
• This approach represents a state-of-the-art advancement in liver tumor semantic segmentation using multi-modal MRI data.