Comparison-Based Convolutional Neural Networks for Cervical Cell/Clumps Detection in the Limited Data Scenario
Automated detection of cervical cancer cells or cell clumps has the potential to significantly reduce error rate and increase productivity in cervical cancer screening. However, most traditional methods rely on the success of accurate cell segmentation and discriminative hand-crafted features extraction. Recently there are emerging deep learning-based methods which train convolutional neural networks to classify image patches, but they are computationally expensive. In this paper we propose to an end-to-end object detection methods for cervical cancer detection. More importantly, we develop the Comparison detector based on Faster-RCNN with Feature Pyramid Network(baseline model) to deal with the limited-data problem. Actually, the key idea is that classify the region proposals by comparising with the prototype representations of each category which learn from reference images. In addition, we propose to learn the prototype representations of the background from the data instead of reference images manually choosing by some heuristic rules. Comparison detector shows significant improvement for small dataset, achieving a mean Average Precision (mAP) 26.3% and an Average Recall (AR) 35.7%, both improving about 20 points compared to baseline model. Moreover, Comparison detector achieves better performance on mAP compared with baseline model when training on the medium dataset, and improves AR by 4.6 points. Our method is promising for the development of automation-assisted cervical cancer screening systems.
- CUDA==9.1
- cuDNN==7.0
- tensorflow==1.8.0
If you want to check the effect, you can download the test set in here and put it under the tfdata/tct. As same time, you must download the weight of model and unzip in the home directory.
We provide evaluate_network.ipynb to verify our results. We also provide predict.ipynb to predict results of a single picture.
The dataset consists of 7410 cervical microscopical images which are cropped from the whole slide images (WSIs) obtained by Pannoramic MIDI II digital slide scanner. In the dataset, there are totally 48,587 instances belonging to 11 categories. We randomly divide the dataset into training set Df which contains 6666 images and test set which contains 744 images. The small training set Ds contains 762 images randomly chosen from Df.
Original image cropped from WSI
Some instances in 11 categories
The dataset is available on Google driver here.