MVT-Net: A novel cervical tumour segmentation using multi-view feature transfer learning

Area of Science:

• Medical Imaging
• Oncology
• Computer Vision
• Machine Learning

Background:

• Cervical cancer is a highly aggressive malignancy threatening women's health globally.
• Accurate segmentation of cervical tumours in MR images is crucial but challenging due to tumour complexity and traditional method limitations.

Purpose of the Study:

• To develop a novel cervical tumour segmentation model, MVT-Net, addressing current segmentation challenges.
• To leverage multi-view feature transfer learning for enhanced tumour characterization in MR images.

Main Methods:

• Proposed MVT-Net integrates a 2D encoder-decoder network (source domain) and a 3D multi-scale segmentation network (target domain).
• Employs a transfer learning strategy to extract diverse, multi-perspective tumour features.
• Incorporates multi-scale residual and attention blocks within the 3D network to capture complex feature correlations.

Main Results:

• MVT-Net achieved superior performance on a 160-image cervical MR dataset compared to state-of-the-art methods.
• Demonstrated high accuracy with a DICE score of [Formula: see text] and an average surface distance (ASD) of [Formula: see text] mm.
• Showcased improved tumour localisation, shape delineation, and edge segmentation accuracy.

Conclusions:

• MVT-Net represents a significant advancement in cervical tumour segmentation technology.
• The multi-view feature transfer learning strategy effectively enhances segmentation accuracy and reliability.
• The model shows promise for improved clinical applications in cervical cancer diagnosis and treatment planning.