My Bachelor’s thesis is about using Deep Learning, on lung X-Ray images, to predict whether the patient has covid or not. I was thinking of using ResNet, but I am at the end of Week 3 where we learn about U-Net, and I am 99% sure, U-Net is a better architecture for my needs.
The datasets structure is Train/Val/Test with 38k images total. Each set has 3 classes. Covid, Non Covid and Normal.
And each class has 2 types of images. Normal X-Ray images and Lung masks.
Normal X-Ray:
Lung Masks:
Any thoughts on using U-Net? Should I look into a different network? Also, so far I have trained the NN in all PAs with images to output a prediction. This time I have images and lung masks, which one of these do I feed into the NN? Both? I haven’t finished week 3 yet, so this question might be answered on later videos or in the PA assignment.
1 Like
Hi, @Marios_Constantinou!
From your post I understand that your objective is solving a classification problem (3 classes) and you have 2 streams of input data for that: x-rays and lung masks. Am I right?
U-Net would be a good choice if your problem was about semantic segmentation, which would be predicting the masks from the x-ray images. However, this is not the case. Maybe using just a normal classification backbone and masking out the irrelevant pixels from your images can be a good way to making the most out of all the available data.
U-Net would be a good choice if your problem was about semantic segmentation, which would be predicting the masks from the x-ray images. However, this is not the case.
Alright, back to the drawing board I guess, haha. Thank you for your input. I will give it a bit more thought and maybe post back here again to see what you think.
Hello, @Marios_Constantinou!
Based on your post, it seems like your goal is to classify images into three categories, and you have two types of data available for this task: X-ray images and corresponding lung masks. If my understanding is correct, then U-Net would be a suitable approach if your objective was to perform semantic segmentation, which involves predicting the masks from the X-ray images. However, since your goal is classification, it may be more effective to use a standard classification algorithm with a masking process that eliminates irrelevant pixels from the images. This method may allow you to utilize all of the available data effectively.
