Pre-processing for Automatic Cervical Cancer Screening: Image Segmentation and Movement Detection
1Signal Processing Laboratory 5 (LTS5), École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland
2Department of Radiology, University Hospital Center (CHUV) and University of Lausanne, Lausanne, Switzerland
3ETSETB, Universitat Politècnica de Catalunya (UPC), Barcelona, Spain
Bachelor's thesis work at LTS5, EPFL: fine-tuned YOLOv5 to automatically crop the cervix region of interest for VIA screening, reaching a 0.96 Dice score robust to blood, mucus, low light and blur.
Abstract
Cervical cancer remains a major burden in settings where access to healthcare is limited. A smartphone-based artificial intelligence tool was developed to automatically detect and localize precancerous cervical lesions, using the pixels' whitening evolution from videos taken during visual inspection with acetic acid. While around 23.5% of the videos could not be used because of strong movement, this approach also required manual segmentation of the region of interest (ROI). The current work automatizes these pre-processing steps.
Our solution uses 120s videos where the cervical ROIs were annotated in the central frames. A neural network was adapted from You Only Look Once (YOLO) and trained with 650 annotated videos. Cross-validation with nine folds and data augmentation is used. As a result, an ROI is detected within each video frame.
Videos containing strong movement have then been manually identified and used to develop a method able to distinguish them. The ROIs of all the video frames are overlapped to generate a heatmap which is automatically processed to quantify the movement.
The segmentation algorithm achieves a Dice coefficient of 0.957 ± 0.004. The ROI is even detected in videos with blood, mucus, low-light, and blurriness. Strong movement is identified with accuracy, precision and recall of 0.980, 0.961 and 1.0, respectively. Integration of this pre-processing pipeline into the classification tool highly reduces the manual pre-processing required.
Talks
Pre-processing for Automatic Cervical Cancer Screening: Image Segmentation and Movement Detection
MICAD 3rd Conference on Medical Imaging and Computer-Aided Diagnosis
