训练具有多个文件夹和子文件夹的 CNN 模型

Question

I am developing a convolution neural network (CNN) model to predict whether a patient in category 1,2,3 or 4. I use Keras on top of TensorFlow.

I have 64 breast cancer patient data, classified into four category (1=no disease, 2= …., 3=….., 4=progressive disease). In each patient's data, I have 3 set of MRI scan images taken at different dates and inside each MRI folder, I have 7 to 8 sub folders containing MRI images in different plane (such as coronal plane/sagittal plane etc).

I learned how to deal with basic “Cat-Dog-CNN-Classifier”, it was easy as I put all the cat & dog images into a single folder to train the network. But how do I tackle the problem in my breast cancer patient data? It has multiple folders and sub-solders.

Please suggest.

Answer 1

You need to build your dataset navigating through folders. This can be done easily in python.

So, suppose you have the following folder tree:

root/
\__ Patient_1/
\__ \_______ MRI_1/
\__ \_______ \___ folder_1/
\__ \_______ \___ folder_2/
\__ \_______ \___ folder_3/
\__ \_______ \___ ...     /
\__ \_______ \___ folder_8/
\__ \_______ MRI_2/
\__ \_______ ...  /...
\__ Patient_2/
...

First, you need to get your current working directory (cwd):

import os
cwd = os.getcwd()

Then you can start building your dataset:

dataset = [];

for p in range(1, 65):
    for mri in range(1, 4):
        mri_folder = cwd + '/Patient_{}/MRI_{}/'.format(p, mri)

        # Since we do not know how many folders there will be, 
        # we need to list all the folders in the current MRI folder.

        folders_names = os.listdir(mri_folder)
        for f in folder_names:
            # list all images in current folder
            fnames = os.listdir('{}/{}/'.format(mri_folder, f))

            for fname in fnames:
                # Now you can open your image and append it to dataset

Hope it helps!

Answer 2

use this loop:

for _,_,files in os.walk("data_folder/"):
    for name in files:
        FP = os.path.join(root, name)

Answer 3

使用os.walk递归访问子目录中的所有文件os.walk加到数据集。