Application of TransUnet Deep Learning Model for Automatic Segmentation of Cervical Cancer in Small-Field T2WI Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of cervical cancer tissue in MRI is vital for diagnosis and treatment.
• Current segmentation methods may lack precision in capturing complex lesion details.

Purpose of the Study:

• To develop and evaluate an innovative deep learning model for enhanced automatic segmentation of cervical cancer lesions in MRI.
• To compare the performance of the proposed model against existing methods.

Main Methods:

• Utilized a dataset of 4063 T2WI MRI images from 222 cervical cancer patients.
• Employed Convolutional Neural Networks (CNNs) for local feature extraction and Transformers for long-range dependencies.
• Developed and assessed three TransUnet models based on coronal, axial, and sagittal views.

Main Results:

• The TransUnet model achieved an average Dice Similarity Coefficient (DSC) of 0.7628 and Mean Hausdorff Distance (AHD) of 0.8687.
• TransUnet outperformed the U-Net model, showing a DSC improvement of 0.0033 and AHD improvement of 0.3479.
• Demonstrated superior segmentation accuracy and efficacy in delineating cervical cancer tissues.

Conclusions:

• The proposed TransUnet model significantly enhances the accuracy of cervical cancer tissue segmentation compared to alternative models.
• This automated image analysis tool shows potential for improving clinical diagnostic efficiency in cervical cancer detection.