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

## TL;DR

This paper introduces MVT-Net, a new model for accurately segmenting cervical tumors in MR images using multi-view feature transfer learning.

## Contribution

The novel MVT-Net model combines 2D and 3D networks with multi-view transfer learning for improved cervical tumor segmentation.

## Key findings

- MVT-Net achieved a DICE score of 75.9±7.43% on cervical MR images.
- The model outperformed existing methods in tumor localization and edge segmentation.
- Ablation studies confirmed the effectiveness of the multi-view feature transfer strategy.

## Abstract

Cervical cancer is one of the most aggressive malignant tumours of the reproductive system, posing a significant global threat to women’s health. Accurately segmenting cervical tumours in MR images remains a challenging task due to the complex characteristics of tumours and the limitations of traditional methods. To address these challenges, this study proposes a novel cervical tumour segmentation model based on multi-view feature transfer learning, named MVT-Net. The model integrates a 2D global axial plane encoder-decoder network and a 3D multi-scale segmentation network as source and target domains, respectively. A transfer learning strategy is employed to extract diverse tumour-related information from multiple perspectives. In addition, a multi-scale residual blocks and a multi-scale residual attention blocks are embedded in the 3D network to effectively capture feature correlations across channels and spatial positions. Experiments on a cervical MR dataset of 160 images show that our proposed MVT-Net outperforms state-of-the-art methods, achieving a DICE score of 75.9±7.43%, an ASD of 2.69±0.58 mm and superior performance in tumour localisation, shape delineation and edge segmentation. Ablation studies further validate the effectiveness of the proposed multi-view feature transfer strategy. These results demonstrate that our proposed MVT-Net represents a significant advance in cervical tumour segmentation, offering improved accuracy and reliability in clinical applications.