Jittor Medical Segmentation Lib -- The assignment of Pattern Recognition course (2021 Spring) in Tsinghua University

Related tags

Deep LearningJMedSeg
Overview

THU模式识别2021春 -- Jittor 医学图像分割

模型列表

本仓库收录了课程作业中同学们采用jittor框架实现的如下模型:

  • UNet
  • SegNet
  • DeepLab V2
  • DANet
  • EANet
  • HarDNet及其改动HarDNet_alter
  • PSPNet
  • OCNet
  • OCRNet
  • DLinkNet
  • AttentionUNet
  • UNet++
  • UNet+++
  • DenseUNet
  • TernausNet
  • CSNet
  • SCSENet
  • U2Net
  • U2Net-small(轻量化的U2Net)
  • Multi-ResUNet
  • R2 UNet
  • R2 Attention UNet
  • LightNet
  • OneNet(轻量化的UNet)
  • CENet
  • LRF-EANet
  • SimpleUNet
  • SETR

课程同学提出的优秀方案

增加小模型的鲁棒性

成员:汪元标,黄翰,李响,郑勋

  • Spatial Transformer Network(STN)

    通过对图片做自适应的仿射变换提高模型的鲁棒性

    其中Localization Network采用了多层卷积block的结构,输出通道数分别为8, 16, 32, 64,每个block包括一个3x3的二维卷积,BatchNorm层以及ReLU激活函数,最后经过一个Adaptive Average Pool层实现对多尺寸的支持,参见 advance/stn.py

  • 对比学习自监督预训练

    为了学习CT图像的Latent Feature,采用了对比学习的算法,使用了InfoNCE作为损失函数,采用Memory Bank来采样负例。

  • 数据增强

    采取了颜色空间的变换,包含亮度、对比度、色相、色调的随机变换。

  • 效果

    经过改动后的UNet有很好的鲁棒性。即使是智能手机拍摄的照片也可以很好地识别

  • demo

    • 参见清华云盘链接

    • 用法

      > export FLASK_APP='app.py'
      > python -m flask run -p [PORT]

轻量化模型

游嘉诚

为了降低参数量,考虑使用组卷积(group conv),然而组卷积限制了通道之间的信息交流。传统channel shuffle 限制了各 group 信息交流的表达力,同时内存访问连续性差或MACS大。于是提出了领域通道平移(channel shift),即通道顺序平移0.5*group,保证各组只与邻域交流,也许可能使得关系密切的组也许会趋向聚在一起。网络开始使用$4\times4$卷积核和stride=4进行四倍下采样,最终使用转置卷积ConvTranspose进行4倍上采样。类似densenet思想,网络将特征图进行通道拼接后进行转置卷积上采样。

二、数据

腰椎骨松质分割数据集

协和医院和中山医院分别提供了腰椎数据集,经过同学们的标注,从协和数据集中划分出了训练集、验证集和测试集。

协和数据集选择并标注了腰椎3,腰椎4的CT片,其中:

  • 训练集:包含85人共计1442张腰椎CT片

  • 验证集:包含31人共计549张腰椎CT片

  • 测试集:包含36人共计615张腰椎CT片

下图展示了协和数据集的样例

中山数据集被用于测试模型的跨数据集泛化性能,中山数据集包含大图和小图两种类型,选择并标注了腰椎1,腰椎2的CT片。其中:

  • 大图:包含134人共计858张腰椎CT片

  • 小图:包含134人共计868张腰椎CT片

下图分别是中山数据集的大图和小图样例

胰腺分割数据集

胰腺分割数据集包含

  • 训练集:包含8人共计1720张CT片

  • 验证集:包含2人共计424张CT片

  • 测试集:包含2人共计401张CT片

下载地址

清华云盘 数据下载成功后将压缩包解压为 ./data 文件夹即可。

运行方法

  1. 配置相应环境,安装Jittor最新版本

  2. 下载数据集

  3. 下载一些模型必须的ImageNet预训练权重,并解压至model/目录下.

  4. 如想要使用训练好的模型参数,可在这里下载胰腺数据上训练完成的模型参数,也可以先下载可视化结果先观察模型效果。

  5. 运行训练/测试/可视化

usage: run.py [-h]
              [--model {unet,hrnet,setr,unet3p,segnet,hardnet,deeplab,pspnet,danet,eanet,ocrnet,resunet,ocnet,attunet,dense,dlink,ternaus,scseunet,r2,r2att,csnet,unetpp,unetppbig,multires,u2net,u2netp,onenet,lightnet,cenet,setr,hardalter,lrfea,simple}]
              [--pretrain] [--checkpoint CHECKPOINT] --dataset
              {xh,xh_hard,zs_big,zs_small,pancreas} --mode
              {train,test,predict,debug} [--load LOAD] [--aug] [--cuda]
              [--stn] [-o {Adam,SGD}] [-e EPOCHS] [-b BATCH_SIZE] [-l LR]
              [-c CLASS_NUM] [--loss LOSS] [-w BCE_WEIGHT] [-r RESULT_DIR]
              [--poly]

optional arguments:
  -h, --help            show this help message and exit
  --model {unet,hrnet,setr,unet3p,segnet,hardnet,deeplab,pspnet,danet,eanet,ocrnet,resunet,ocnet,attunet,dense,dlink,ternaus,scseunet,r2,r2att,csnet,unetpp,unetppbig,multires,u2net,u2netp,onenet,lightnet,cenet,setr,hardalter,lrfea,simple}
                        choose the model
  --pretrain            whether to use pretrained weights
  --checkpoint CHECKPOINT
                        the location of the pretrained weights
  --dataset {xh,xh_hard,zs_big,zs_small,pancreas}
                        choose a dataset
  --mode {train,test,predict,debug}
                        select a mode
  --load LOAD           the location of the model weights for testing
  --aug                 whether to use color augmentation
  --cuda                whether to use CUDA acceleration
  --stn                 whether to use spatial transformer network
  -o {Adam,SGD}, --optimizer {Adam,SGD}
                        select an optimizer
  -e EPOCHS, --epochs EPOCHS
                        num of training epochs
  -b BATCH_SIZE, --batch-size BATCH_SIZE
                        batch size for training
  -l LR, --learning-rate LR
                        learning rate
  -c CLASS_NUM, --class-num CLASS_NUM
                        pixel-wise classes
  --loss LOSS           Choose from 'ce', 'iou', 'dice', 'focal', if CE loss
                        is selected, you should use a `weight` parameter
  -w BCE_WEIGHT         use this weight if BCE loss is selected; if w is
                        given, then the weights for positive and negative
                        classes will be w and 2.0 - w respectively
  -r RESULT_DIR, --resultdir RESULT_DIR
                        test result output directory
  --poly                whether to use polynomial learning rate scheduler
  1. 运行对比学习预训练
usage: run_ssl.py [-h]
                  [--model {unet,hrnet,setr,unet3p,segnet,hardnet,deeplab,pspnet,danet,eanet,ocrnet,resunet,ocnet,attunet,dense,dlink,ternaus,scseunet,r2,r2att,csnet,unetpp,unetppbig,multires,u2net,u2netp,onenet,lightnet,cenet,setr,hardalter,lrfea,simple}]
                  --dataset {xh,xh_hard,zs_big,zs_small,pancreas}
                  [--save SAVE] [-e EPOCHS] [-c CLASS_NUM] [-b BATCH_SIZE]
                  [--channel EMBEDDING_CHANNEL] [--layer LAYER] [--lr LR]
                  [--pretrain]

optional arguments:
  -h, --help            show this help message and exit
  --model {unet,hrnet,setr,unet3p,segnet,hardnet,deeplab,pspnet,danet,eanet,ocrnet,resunet,ocnet,attunet,dense,dlink,ternaus,scseunet,r2,r2att,csnet,unetpp,unetppbig,multires,u2net,u2netp,onenet,lightnet,cenet,setr,hardalter,lrfea,simple}
                        choose a model network
  --dataset {xh,xh_hard,zs_big,zs_small,pancreas}
                        select a dataset
  --save SAVE           model weights save path
  -e EPOCHS, --epochs EPOCHS
                        number of training epochs
  -c CLASS_NUM, --class-num CLASS_NUM
                        class number
  -b BATCH_SIZE, --batch-size BATCH_SIZE
                        training batch size
  --channel EMBEDDING_CHANNEL
                        number of channels of embedded feature maps
  --layer LAYER         layer to extract features from
  --lr LR               learning rate
  --pretrain

运行示例 参见train.sh, batch_test.sh, pretrain.sh

四、实验结果

胰腺分割数据集

各模型均采用相同超参数,学习率为3e-4,迭代次数50次,以权重为[0.8, 0.2]的交叉熵损失函数进行训练,参见train.sh中dataset为pancreas的部分。结果如下:

Model Dice mIoU
UNet 0.7292 0.6477
SegNet 0.6291 0.5726
DeepLab 0.8306 0.7467
DANet 0.7787 0.6928
EANet 0.6753 0.6055
HarDNet 0.7491 0.6654
HarDNet_alter 0.7779 0.6920
PSPNet 0.7772 0.6914
OCNet 0.7789 0.6930
OCRNet 0.7034 0.6272
DLinkNet 0.4995 0.4989
AttentionUNet 0.7691 0.6836
UNet++ 0.8282 0.7439
UNet+++ 0.7892 0.7030
DenseUNet 0.8053 0.7194
TernausNet 0.6752 0.6055
CSNet 0.4994 0.4989
SCSENet 0.4994 0.4989
U2Net 0.8143 0.7289
U2Net-Small 0.8338 0.7502
Multi-ResUnet 0.7230 0.6427
R2UNet 0.8289 0.7447
R2AttentionUNet 0.8084 0.7227
LightNet 0.8006 0.7145
OneNet 0.7754 0.6896
CENet 0.7583 0.6735
LRF-EANet 0.6942 0.6197
SimpleUNet 0.7395 0.6569
SETR 0.4994 0.4989
UNet-SSL 0.8026 0.7165
UNet-STN-SSL 0.7926 0.7063
UNet-Aug-STN-SSL 0.6938 0.6192

腰椎骨松质分割数据集

  1. 协和数据集

    下表展示了三种模型使用不同损失函数及加权组合的结果

    mIoU CE IoU Dice Focal 0.8CE+0.2IoU 0.5CE+0.5IoU 0.2CE+0.8IoU 0.5CE+0.5Dice
    UNet 95.49 95.56 95.41 95.43 95.48 95.25 95.42 95.17
    HRnet 95.22 95.23 N/A N/A
    SETR 87.59 87.81 85.92 83.85 88.34 83.94 87.52 87.78
  2. 中山数据集

    下表记录了协和数据集上训练的OneNet模型在中山数据集上的表现

Reference code for the paper CAMS: Color-Aware Multi-Style Transfer.

CAMS: Color-Aware Multi-Style Transfer Mahmoud Afifi1, Abdullah Abuolaim*1, Mostafa Hussien*2, Marcus A. Brubaker1, Michael S. Brown1 1York University

Mahmoud Afifi 36 Dec 04, 2022
An improvement of FasterGICP: Acceptance-rejection Sampling based 3D Lidar Odometry

fasterGICP This package is an improvement of fast_gicp Please cite our paper if possible. W. Jikai, M. Xu, F. Farzin, D. Dai and Z. Chen, "FasterGICP:

79 Dec 31, 2022
Python script to download the celebA-HQ dataset from google drive

download-celebA-HQ Python script to download and create the celebA-HQ dataset. WARNING from the author. I believe this script is broken since a few mo

133 Dec 21, 2022
This repo is to present various code demos on how to use our Graph4NLP library.

Deep Learning on Graphs for Natural Language Processing Demo The repository contains code examples for DLG4NLP tutorials at NAACL 2021, SIGIR 2021, KD

Graph4AI 143 Dec 23, 2022
Code for the paper "Multi-task problems are not multi-objective"

Multi-Task problems are not multi-objective This is the code for the paper "Multi-Task problems are not multi-objective" in which we show that the com

Michael Ruchte 5 Aug 19, 2022
Solver for Large-Scale Rank-One Semidefinite Relaxations

STRIDE: spectrahedral proximal gradient descent along vertices A Solver for Large-Scale Rank-One Semidefinite Relaxations About STRIDE is designed for

48 Dec 20, 2022
An Implementation of Fully Convolutional Networks in Tensorflow.

Update An example on how to integrate this code into your own semantic segmentation pipeline can be found in my KittiSeg project repository. tensorflo

Marvin Teichmann 1.1k Dec 12, 2022
The code used for the free [email protected] Webinar series on Reinforcement Learning in Finance

Reinforcement Learning in Finance [email protected] Webinar This repository provides the code f

Yves Hilpisch 62 Dec 22, 2022
Breaking the Curse of Space Explosion: Towards Efficient NAS with Curriculum Search

Breaking the Curse of Space Explosion: Towards Effcient NAS with Curriculum Search Pytorch implementation for "Breaking the Curse of Space Explosion:

guoyong 17 Jan 03, 2023
NEATEST: Evolving Neural Networks Through Augmenting Topologies with Evolution Strategy Training

NEATEST: Evolving Neural Networks Through Augmenting Topologies with Evolution Strategy Training

Göktuğ Karakaşlı 16 Dec 05, 2022
Continual World is a benchmark for continual reinforcement learning

Continual World Continual World is a benchmark for continual reinforcement learning. It contains realistic robotic tasks which come from MetaWorld. Th

41 Dec 24, 2022
public repo for ESTER dataset and modeling (EMNLP'21)

Project / Paper Introduction This is the project repo for our EMNLP'21 paper: https://arxiv.org/abs/2104.08350 Here, we provide brief descriptions of

PlusLab 19 Oct 27, 2022
CRNN With PyTorch

CRNN-PyTorch Implementation of https://arxiv.org/abs/1507.05717

Vadim 4 Sep 01, 2022
Nicely is a real-time Feedback and Intervention Program Depression is a prevalent issue across all age groups, socioeconomic classes, and cultural identities.

Nicely is a real-time Feedback and Intervention Program Depression is a prevalent issue across all age groups, socioeconomic classes, and cultural identities.

1 Jan 16, 2022
This is an early in-development version of training CLIP models with hivemind.

A transformer that does not hog your GPU memory This is an early in-development codebase: if you want a stable and documented hivemind codebase, look

<a href=[email protected]"> 4 Nov 06, 2022
An implementation for the loss function proposed in Decoupled Contrastive Loss paper.

Decoupled-Contrastive-Learning This repository is an implementation for the loss function proposed in Decoupled Contrastive Loss paper. Requirements P

Ramin Nakhli 71 Dec 04, 2022
LEAP: Learning Articulated Occupancy of People

LEAP: Learning Articulated Occupancy of People Paper | Video | Project Page This is the official implementation of the CVPR 2021 submission LEAP: Lear

Neural Bodies 60 Nov 18, 2022
Parallel and High-Fidelity Text-to-Lip Generation; AAAI 2022 ; Official code

Parallel and High-Fidelity Text-to-Lip Generation This repository is the official PyTorch implementation of our AAAI-2022 paper, in which we propose P

Zhying 77 Dec 21, 2022
A generalist algorithm for cell and nucleus segmentation.

Cellpose | A generalist algorithm for cell and nucleus segmentation. Cellpose was written by Carsen Stringer and Marius Pachitariu. To learn about Cel

MouseLand 733 Dec 29, 2022
Keras implementation of "One pixel attack for fooling deep neural networks" using differential evolution on Cifar10 and ImageNet

One Pixel Attack How simple is it to cause a deep neural network to misclassify an image if an attacker is only allowed to modify the color of one pix

Dan Kondratyuk 1.2k Dec 26, 2022