Can this project not be trained except for data files in *.mhd format?

在使用您的代码时，我设置2d图像训练，在main.py文件中255行和277行，应该怎么修改吗？

Hi there,

I found your work very inspiring, and I made my own segmentation model follwing your idea. But I found a problem in the validation step.

In your main.py file, line 456:

for i,subj in enumerate(test_dataset.subjects):
...

I think in the enumerate parentheses, it should not be test_dataset.subjects because the the subjects object is the raw image list without preprocessing (aka transforms in the data_function.py). Maybe here should be:

for i,subj in enumerate(test_dataset.valid_set):
...

Because valid_set(also there may be a typo that you use training_set in test_loader) is the one with transforms(normalization) applied.

Maybe my comprehension is not correct and you have your own thoughts, so could you please share your idea on this step? Thank u

Do you have pre-trained checkpoints for releasing??

How can we inference on test data without having labels?

新入门医学影像分割小白，请问您有医学影像目标分割的数据集吗，在网上找了一段时间，但是都没有找到合适的香瓜的数据集，如果方便的话希望分享以下，谢谢🙏

请问一下，图片可以直接使用png或者jpeg的图片或者标注吗？

请问一下你的数据集是怎么划分的，训练集、验证集、测试集？

Hello, dear author!
Thank you for your amazing job!
Could you tell me how to train 2D images with multi-classes (assume num_classes=5)?
Where should I modify? (expect "mode = '2d', in_class = 3, out_class = 5" in hparam.py)
Thank you in advance!

作者你好，首先很感谢您在github上提供的代码。
下面是我在训练3D CT数据出现的问题：

训练时在运行到main.py里的260行，也就是loss=criterion_ce(outputs,y.argmax(dim=1)) +criterion_dice(outputs,y.argmax(dim=1))出错，我进去看后，发现是predict和target维度不一致，以图片大小为646496，batchs=4，channel=2为例，predict为（4,2,64,64,96），target为（4,128,64,96），缺少channel维度，使得dice损失函数无法运行，同时self._one_hot_encoder函数使得第二个数翻倍，也就是叠加了，我加入target = torch.unsqueeze(target,1)这行代码后，成功进行训练，但tensorboard中显示loss在缓慢下降，但dice直接下降到0。

D:\ProgramData\miniconda3\envs\pytorch\python.exe E:/python_learning/Pytorch-Medical-Segmentation-master/main.py
epoch:1
Batch: 0/5 epoch 1

Traceback (most recent call last):
File "E:\python_learning\Pytorch-Medical-Segmentation-master\main.py", line 582, in
train()
File "E:\python_learning\Pytorch-Medical-Segmentation-master\main.py", line 265, in train
loss = criterion_ce(outputs, y.argmax(dim=1)) + criterion_dice(outputs, y.argmax(dim=1))
File "D:\ProgramData\miniconda3\envs\pytorch\lib\site-packages\torch\nn\modules\module.py", line 1194, in _call_impl
return forward_call(*input, **kwargs)
File "E:\python_learning\Pytorch-Medical-Segmentation-master\loss_function.py", line 141, in forward
assert inputs.size() == target.size(), 'predict & target shape do not match'
AssertionError: predict & target shape do not match

Process finished with exit code 1

非常感谢您的回复

肺结节的标注是写在csv中的，请问有什么方法能够加载它作为Label吗

想请教一下这个操作是基于什么样的理念要求的？

I trained correctly, but when test, this error occurs, I don't know how to fix it?

现在的模型训练后只是把脊椎的所有节都分成一个颜色，我的训练集标签是每相邻节骨头都是不一样的颜色，我也需要这样的推理效果，请问应该怎么处理啊？

我用您更新后的代码重新训练了模型，在预测时出现了这样的错误：
File "main.py", line 519, in test
aggregator_1.add_batch(labels, locations)
File "/home/anran/.conda/envs/seg/lib/python3.6/site-packages/torchio/data/inference/aggregator.py", line 131, in add_batch
self.patch_overlap,
File "/home/anran/.conda/envs/seg/lib/python3.6/site-packages/torchio/data/inference/aggregator.py", line 82, in crop_batch
diff = patch_shape - crop_shape
ValueError: operands could not be broadcast together with shapes (2,) (3,)

请问要怎么解决

泪目。。。。

RuntimeError: More than one value for "spacing" found in subject images:
{'label': (1.0468800067901611, 1.0468800067901611, 2.5),
'source': (0.712891, 0.712891, 2.49998)}

你好，我的数据是3d的，后缀nii.gz,请问metrics.py加载数据目录的构成应该是什么样的？
应该没问题，我是在这一行出错的，for i,subj in enumerate(training_set.subjects):
报错为：
File "/home/carrie/PycharmProjects/Pytorch_Medical_Segmentation/metrics.py", line 55, in
for i,subj in enumerate(training_set.subjects):#training_set.subjects
AttributeError: 'SubjectsDataset' object has no attribute 'subjects'
可以再readme里加一个metrics运行示例吗？
可以了，这样改 for i,subj in enumerate(training_set._subjects):#少了一个下划线

但是我在代码层面修改了无法读取不同类型文件的问题，但是训练出来的结果很差。。。，会不会是因为这个问题导致的

Do you have pre-trained checkpoints for releasing??

你好，很感谢您的代码，我预测效果很差，多次修改损失函数都没有效果，看代码发现这里有点疑问，请问这里是不是错了？
man.py代码:
logits = torch.sigmoid(outputs)
labels = logits.clone()
labels[labels>0.5] = 1
labels[labels<=0.5] = 0
loss = criterion(outputs, y)#有问题？应该为loss = criterion(logits, y)

你好，想请问下如何设置diceloss里面的weight。里面的weight是不是充当weight_loss 的作用？

I fixed the errors in the model prediction in the same way as you, but now the prediction results are all black. What's the problem. What I did was a dichotomy. During the test, there was such a warning:

我自己3d脊椎训练的数据集是7张nii.gz文件，epoch为20，batch为4.
设置aug=true后，进行推理时，手动输出outputs最大值小于0.5，导致result图像为空，但是设置false后就会正常（指result最大值会大于0.5）。
目前还没试验过其他数据集，这一点可以注意一下。

Can you provide pretrained model of 3D segmentation ?

If Yes? How to use? Thanks.

How to solve this problem?
input size:torch.Size([2, 2, 32, 32, 32]) target size: torch.Size([2, 64, 32, 32])

你好，我分割训练LITS肝脏，tensorboard可视化，loss，dice我觉得都还不错。


左边是预测的，右边是分割的，但是我测试的结果可视化之后，一点都没分割出来，我找了好久也没发现原因，想问问您，是拼接的问题吗？？

我使用这个版本的 ResUNet 在我的数据上进行训练可以得到0.78左右的 Dice 值. 我使用您的模型对同样的数据进行训练模型无法进行收敛 Dice 值始终处在0.1以下. 我在对模型进行检查后仍未能发现问题.

这是我训练的 log. 其中红色线为2DResUNet, 蓝色线为3D模型.

我原始的数据类型为 .dcm 类型, 使用脚本转为 .nii.gz 格式. 我使用 dicompyler 对转后的 .nii.gz 进行检查未发现数据存在问题.

希望能够得到您的回复!

使用3d unet，用的brats2020的数据集，python main.py后，targetsize和inputsize会出现维度不匹配的问题，inputsize是[(2,2,32,32,32)],targetsize是[(2,64,32,32)]；我把target.reshape([2,2,32,32,32])后成功运行，但是后续调整batsize的的大小，程序又会报错，RuntimeError: shape '[2, 2, 32, 32, 32]' is invalid for input of size 524288，就是inputsize会随着batchsize的改变而改变，这是什么原因呢，因为一般batchsize不就是可以随意改变的吗

Thanks for your github.

I'm testing with .nrrd format.
however, there is an issue where the resulting image is enlarged.
(ex: input size:256x256x256 --> result : 128x128x128)
Also, the image is enlarged to 32-bit or greater precision.
What part of main.py should I edit?

您好，我看了您的新论文“A learnable Gabor Convolution kernel for vessel segmentation”，请问可以分享这篇论文的代码嘛？万分感谢！

Is data and pretrained model openly available? Please share hyperlinks.

您好！有一个关于训练结果的问题想要请教一下。
我使用的网上下载的胰腺数据集，原数据和标签都是 nii.gz 格式的，因为是3d数据，所以用的是3d的网络来跑的。是在本地的计算机上跑的代码。默认设置的epoch是100
参数设置如下：

但是我训练到后面，loss值是在下降，但是dice值就一直是0了，false_positive_rate以及false_negative_rate也很不正常，如下所示：

想请问原因可以能为什么呢？是否和epoch数有关？因为我看您的原代码是500000，我设置100已经需要跑一段时间了。
其他的参数是否有要根据数据的三维尺寸调整的呢？
感谢回复！

can gave an example ,thanks

我看了下可能因为数据量比较大 一个epoch 就收敛的差不多了 但是我查看TensorBoard显示dice false_negtive_rate false_positive_rate 都没结果 是我的姿势不对嘛。

这该怎么办？

Hi MontaEllis team!

The Medical Segmentation project is fantastic. I see you host and share models with your server. Would you be interested in sharing your models in the Hugging Face Hub?

This integration would allow you to freely download/upload models and make your work more accessible and visible to the rest of the ML community. We can help you set up a MontaEllis organization (examples, Facebook AI y Stanford NLP).

Creating the repos and adding new models should be a relatively straightforward process. This is a step-by-step guide explaining the process in case you're interested. Please let us know if you would be interested and if you have any questions.

Some of the benefits of sharing your models through the Hub would be:

• Presence in the HF Hub might lower the entry of barrier to your medical models and increase their visibility.
  • Repos provide valuable metadata about their tasks, languages, metrics, etc. that make them discoverable
• versioning, commit history, and diffs.
• multiple features from TensorBoard visualizations, PapersWithCode integration, and more.

Additionally, we have a library to access repositories programmatically (both downloading pre-trained models and pushing, with many nice things such as filtering, caching, etc.). If we want to try out this integration, I would suggest you add one or two models manually and then use the huggingface_hub library to implement downloading those models programmatically from pytorch-medical-segmentation. To read more about it, you might want to check our documentation to read more about it.

Relevant references:

• Adding a model to the Hub.
• Integrate your library with the Hub.

Happy to hear your thoughts,

Omar and the Hugging Face team (cc @osanseviero @abidlabs )

你好 请问在哪里添加预处理的代码 我没找到您预处理的代码。