SwinUNeCCt: bidirectional hash-based agent transformer for cervical cancer MRI image multi-task learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Cervical cancer is a leading global health concern for women, with significant mortality.
• Magnetic resonance imaging (MRI) is crucial for cervical cancer diagnosis and staging.
• Manual segmentation of cervical cancer in MRI is labor-intensive and prone to subjectivity.

Purpose of the Study:

• To develop an automated segmentation model for accurate identification of cervical cancer lesions in MRI scans.
• To introduce the novel SwinUNeCCt computer-aided diagnosis model.
• To evaluate the performance of SwinUNeCCt against existing state-of-the-art 3D medical imaging models.

Main Methods:

• Utilized a dataset of 122 cervical cancer patients' pelvic dynamic contrast-enhanced MRI scans.
• Developed SwinUNeCCt incorporating a bidirectional hash-based agent multi-head self-attention mechanism.
• Compared SwinUNeCCt with models like nnUnet, TransBTS, and SwinUNetR on semantic segmentation tasks.

Main Results:

• SwinUNeCCt achieved superior performance in semantic segmentation without a classification module, with best-in-class metrics (95HD: 6.25, IoU: 0.669, DSC: 0.802).
• The model demonstrated a strong balance between computational efficiency (442.7 GFLOPs) and model complexity (71.2 M parameters).
• SwinUNeCCt also showed high recognition capabilities in semantic segmentation tasks with an included classification module, outperforming comparative models.

Conclusions:

• The SwinUNeCCt model represents a significant advancement in automated cervical cancer lesion segmentation from MRI.
• It offers improved accuracy and efficiency compared to current state-of-the-art methods.
• The model's balanced computational demands and high performance make it a promising tool for clinical applications in cervical cancer diagnosis.