Unsupervised Deep Learning Registration of Uterine Cervix Sequence Images

Area of Science:

• Medical Imaging
• Computer Vision
• Oncology

Background:

• Colposcopy for cervical cancer prevention involves acquiring digital images of the cervix.
• Dynamic pixel intensity variations during the aceto-whitening reaction offer informative insights.
• Spatial misalignment in image sequences due to patient movement hinders automated visual evaluation (AVE).

Purpose of the Study:

• To develop a novel unsupervised registration approach for aligning sequences of digital cervix color images.
• To address the challenge of lacking registration ground truth for training supervised algorithms.
• To improve the accuracy of disease prediction in cervical cancer screening using AVE.

Main Methods:

• A deep-learning-based registration network with three branches processing RGB channels separately.
• Utilizing convolutional neural network (CNN) units and spatial transform units within each branch.
• An unsupervised strategy for image registration without requiring ground truth data.

Main Results:

• The proposed method demonstrated significant improvement in image registration, with an average Dice score increase of 12.62%.
• Achieved higher Dice and IoU scores compared to a non-deep learning registration method.
• Maintained full image integrity throughout the registration process.

Conclusions:

• The novel unsupervised deep learning approach effectively aligns sequential cervix images, overcoming the lack of ground truth data.
• Accurate image registration is vital for enhancing the performance of automated visual evaluation in cervical cancer screening.
• This method offers a promising solution for improving the reliability and accuracy of AVE in colposcopy.