A Deep Clustering Method For Analyzing Uterine Cervix Images Across Imaging Devices

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Visual inspection of the cervix with acetic acid (VIA) is a traditional method for cervical cancer screening but is prone to errors.
• Automated visual evaluation (AVE) utilizes deep learning on digital images of the acetowhitened cervix to predict precancer, offering a potentially low-cost improvement over human performance.

Purpose of the Study:

• To investigate if images captured by three distinct devices (smartphone, custom handheld, clinical colposcope) can be differentiated based on cervical visual appearance using a deep learning clustering approach.
• To address the challenge of device variability in automated visual evaluation (AVE) for cervical cancer screening.

Main Methods:

• A novel deep learning-based clustering approach was developed, comprising cervix region detection, feature extraction, feature encoding, and clustering.
• The method was applied to images acquired from a common smartphone, a custom smartphone-based handheld device, and a clinical colposcope.

Main Results:

• The proposed deep learning clustering method achieved a high accuracy of 97% in distinguishing images from the three different devices.
• The method significantly outperformed several representative deep clustering techniques on the dataset.
• High clustering performance indicates significant visual appearance disparities in cervix images across devices.

Conclusions:

• The study demonstrates that visual appearance of the cervix varies significantly across different imaging devices, posing a challenge for robust AVE.
• There is a clear need for methods that minimize inter-device image variance to improve AVE generalization.
• Developing device-independent AVE requires extensive training datasets from diverse sources to ensure worldwide applicability.