Thanks to Jianwei Yang and Jiasen Lu, I render their pytorch version of faster-RCNN for bladder detection and got excellent results.
@article{jjfaster2rcnn, Author = {Jianwei Yang and Jiasen Lu and Dhruv Batra and Devi Parikh}, Title = {A Faster Pytorch Implementation of Faster R-CNN}, Journal = {https://github.com/jwyang/faster-rcnn.pytorch}, Year = {2017} }
@inproceedings{renNIPS15fasterrcnn,
Author = {Shaoqing Ren and Kaiming He and Ross Girshick and Jian Sun},
Title = {Faster {R-CNN}: Towards Real-Time Object Detection
with Region Proposal Networks},
Booktitle = {Advances in Neural Information Processing Systems ({NIPS})},
Year = {2015}
}
Some detection Results:
CT/MRI images are collected from TCIA-BLCA (The Cancer Imaging Archive), you can download original data from: https://wiki.cancerimagingarchive.net/display/Public/TCGA-BLCA
These data include:
- CT/MRI Series
- CSV files regarding patient status and staging info where patientID is an important key I use for naming preprocessing images.
Besides, I upload detect-and-crop bladder images and the GAN label here: BaiduNetdisk Link：
https://pan.baidu.com/s/1zOp-11bdr3Tbl3W5yjAiUw
Captcha Code ：jz7h
- You can try download the cropped image first, and try to run code from dcgan folder for T0/T1/T4 images, except that u need small changes to paths variables.
- After you finish all the augmentation works, try training by running:
- python main.py
Also, some changes has to be made like path variable and cuda device etc..
First, I choose useful slices from an CT serie where bladder is presented. But then I found some class dataset are insufficient. There are only 8 sample from T0! So I generated GAN samples for T0, T1 and T4, and merge sub-classes for T2 and T3 in order to balance the dataset, here's the comparison between original and generated dataset sub-classes samples count:
Classify stages of bladder cancer, here I just use fine-tunning + pretrained model for classification.
ROC and AUC of individual class are calculated, beside recall, precision and F1 showed as report
