Brief Introduction

Cervical cancer is the most common malignant tumor in women. A colposcopy is used to find cancerous cells or abnormal cells that can become cancerous in the cervix, vagina, or vulva. One may be referred for a colposcopy if she has had an abnormal screening result, which could be either a persistent low grade abnormality or a high-grade abnormality. Therefore, the accurate diagnosis of colposcopic images is essential for the early detection of cancer.

To alleviate the shortage of medical resources in reading colposcopic images, we develop a semantic segmentation method to help the diagnosis process. The method segments colposcopic images into four categories, namely background (BG), low-grade squamous intraepithelial lesion (LSIL), how-grade squamous intraepithelial lesion (HSIL) and cancer (CA). We work closely with Obstetrics and Gynecology Hospital of Fudan University, and a colposcopic image dataset with more than 10,000 images is established. We adopt semantic FPN as the segmentation network and ResNet-50 pretrained on ImageNet as the backbone. Under the three folds cross validation, the average Dice and Recall of our method are 51.8% and 71.6% on the dataset.

Implementation Details

We make some efforts to meet the challenges in this task and further improve the model performance.

Firstly, the imaging quality of some images is poor, and we develop a aharpness evaluation scheme based on image gradient analysis to filter out such images.

Secondly, the pixels in the foreground and background areas are unbalanced, and we devise a training strategy based on pixel balance region resampling.

Thirdly, the scale of lesions varies greatly. To solve this problem, the training strategy based on multi-scale data augmentation and the feature pyramid network structure are adopted.

Future Works

Some annotations of colposcopic images are inaccurate, which may decrease the model performance. To solve this problem, on the one hand, a new dataset with more accurate annotations needs to be established, and on the other hand, a more suitable training strategy is expected to alleviate the influence of the inaccurate annotations.