Is there any feasible method to automated labelling images for a deep learning task, which I have a lot of images to label which is not practically feasible to do manually

I am doing a medical imaging project with MRI images. I hope to develop a deep learning model for this task. But the data I collected is not labelled and I have to label them for each category. Can someone suggest me a method to automate this labelling process where it will be greatly save my time. Thanks in advance. !

[Kalhara]

@Ganesh

I understand what you are trying to say. So far I have only the labelled set of images which I hope to use for training purposes,(Ground truth) and do not have any algorithmic definition for it. External labelling doesn't guarantee the privacy of medical data right? and somewhat expensive. In order to employ the technique you mentioned above, is it necessarily need an algorithm? Is it so, are there any specific algorithm or a custom algorithm that we should build in our own?

[Ganesh]

@Kalhara

You're right, external labelling doesn't guarantee the privacy of your data.

What I mean is, let's say you have a figure, and one is greyscale one, other is coloured. Now, it's easy for you to categorize these two different pictures. Similarly, if you have two different images, and certain pixels give you information about your data, you can use them to label. Maybe your image has a label in the image itself? Would it be possible to apply an "OCR" and extract it?

Another method I thought might be worth a try is to perhaps first make a model by manually labelling a small set of data. Allow it to make predictions, and then validate them, again, manually, and expand the dataset so you can make a model that has better accuracy. You can use the Image Recognition to employ the same.

[Kalhara]

@Ganesh Your active support in my issue greatly appreciated. I am trying the method but there was some problem in the custom algorithm. I think I have to check it twice. Thank you so much.

Answer 1

[Ganesh]

Hi @Kalhara,

The feasibility completely depends on what labels you intend to provide. If you have an "algorithmic" definition of what each label would correspond to in an image, it's straight forward to achieve this. If you are not able to define it, then you certainly need to take help of another AI/DL Model to help you.

If you have enough image examples for each label, you can build your own DL Classification Model and make predictions. As a final step, you can go through the classifications to ensure that the predictions are accurate.

Another option for you is to use external AI services, and ask an AI Model to classify it. This will cost you externally and your data will not be kept intact with MathWorks alone as you need to send your Image Data complying with the requirements of their service.

Answer 2

[Philip Brown]

@Kalhara, you could take a look at the Medical Image Labeler app from R2022b. That app supports labelling medical images using built-in DL models from the Medical Open Network for AI (MONAI) Label platform. You could alternatively train your own custom DL algorithm and use it to automate labelling in the Medical Image Labeler.

[Kalhara]

@Philip Brown

Thank you for your comment. I got to know about it later, I will definetly try that tool. Your support is greatly appreciated.