DiahannWu commented 2 years ago

大神您好，我想请教一下.npy文件中texture中的1对应非实性/磨玻璃，2对应非实性/混合，3对应部分实性/混合，4对应实性/混合，5对应实性吗，malignancy中1对应极不可能，2对应不太可能，3对应不确定，4对应中度可疑，5对应高度可疑吗？ luna2022分类生成训练集和标签的代码方便分享一下吗？分类的话是不是不需要用Mask呀？如果方便的话能麻烦您写一个luna2022数据集分类的代码执行步骤吗，我是初学者，根据您在GitHub上的代码感觉很难理解

junqiangchen commented 2 years ago

@DiahannWu 您好，luna2022数据分类标签生成可以参考这个链接里的实现https://zenodo.org/record/6559584#.Yv3dJSdBxD8 具体如下：

LUNA 22 prequel

This challenge contains a public training dataset of lung nodules extracted from the LIDC-IDRI dataset. This dataset only contains 3D patches of nodules (size 128x128x64 in x,y,z directions) that had been annotated by at least 3 out of 4 radiologists (following LUNA16 criteria, which produces 1186 nodules). The nodules are always located at the center of the 3D patch. 10 nodules have been excluded as part of a hidden test set for phase 1.

To collect the labels for each nodules, you will have to read the LIDC-IDRI_1176.npy file. An example snippet is below:

import numpy as np

dataset = np.load("LIDC-IDRI_1176.npy", allow_pickle=True)

for i in range(len(dataset)):
    nodule = dataset[i]
    print(nodule)
    break

This produces the following output:

{
    'SeriesInstanceUID': '1.3.6.1.4.1.14519.5.2.1.6279.6001.100225287222365663678666836860',
    'VoxelCoordX': 45,
    'VoxelCoordY': 211,
    'VoxelCoordZ': 77,
    'Diameter': [6.97167141, 6.97167141, 7.34878692, 5.94228451],
    'Texture': [5, 5, 5, 5],
    'Malignancy': [4, 2, 4, 2],
    'Calcification': [6, 6, 6, 6],
    'Filename': '1.3.6.1.4.1.14519.5.2.1.6279.6001.100225287222365663678666836860_45_211_77_0000.nii.gz'
}

You can likewise loop through the entire list of 1176 nodules to find the information in the fields.

You can find additional information for some of the fields below:

Nodule Type (Texture)

int, range = {1,2,3,4,5} – Radiographic solidity: internal texture (solid, ground glass, or mixed).

1. Non-Solid/GGO
2. Non-Solid/Mixed
3. Part Solid/Mixed
4. Solid/Mixed
5. Solid

You may want to reduce these into 3 classes: 1. Non-solid/ GGO, 2. (Non-solid/Mixed, Part-solid/Mixed, Solid/Mixed), and 3. Solid

Malignancy

int, range = {1,2,3,4,5} – Subjective assessment of the likelihood of malignancy, 
assuming the scan originated from a 60-year-old male smoker.

1. Highly Unlikely
2. Moderately Unlikely
3. Indeterminate
4. Moderately Suspicious
5. Highly Suspicious

You may want to combine the classes like in the previous section. Perhaps combine classes 1. and 2. as benign, and 3., 4., and 5. as malignant into a binary classification problem.

NOTE: You will most definitely have to combine the classes and convert them into a binary classification or a multi-class classification problem by combining the labels provided by each radiologist (using a median score for each field is recommended)

DiahannWu commented 2 years ago

您发送的邮件我已经收到，谢谢

junqiangchen / PytorchDeepLearing

代码请教 #2

LUNA 22 prequel

Nodule Type (Texture)

Malignancy