I’m writing to clarify the normalization strategies used across different assignments in AI for Medical Diagnosis (Coursera).
In the C1_W1_Assignment (CT scan multiclass classification), we standardized data using the mean and standard deviation computed from the training set and applied them to the validation and test sets—a standard global normalization approach.
However, in the C3_W3_Assignment (3D MRI segmentation), we used sample-wise normalization, where each image or volume is independently normalized using its own mean and standard deviation across train/val/test sets.I’m curious about the reasoning behind this choice.
From what I’ve researched, CT scans use standardized intensity units (e.g., Hounsfield units), which makes global normalization appropriate. In contrast, MRI intensities vary significantly due to acquisition protocols and hardware differences, making sample-wise normalization more suitable.
Could you kindly confirm if this interpretation is correct?
That’s really good question and trying to find a solution to your query.
Now come to your understanding of c1_w1 approach of normalisation is totally correct but let me tell you that is not a mandatory approach, one could also try the sample wise approach as well but if you notice the assignment is more focused on using data augmentation techniques to increasing the data size after the split, hence the global approach was chosen(it is not mandatory). Also c1 w1, is using chest x-ray images not CT scan images (just to notify your confusion)
For c3_w3, I don’t know if the intensity units are the reason but it helps to standardize the input data, making the model’s interpretation more reliable and consistent. By normalizing each 3D MRI volume individually, you account for variations in image intensity and contrast across different scans, which can be caused by differences in acquisition parameters or subject characteristics. This ensures that the Grad-CAM visualizations focus on meaningful features rather than being influenced by irrelevant variations in overall image brightness or contrast.
@Deepti_Prasad Thanks for your valuable input. This helps clarifying a few things. Meanwhile, I will wait for additional resource you can send my way if available
No resources but it is understanding of medical data in application of Grad-CAM technique.
You probably would find 3d mri related projects on kaggle.
Because Grad-CAM has a complex model with different layers and 3dMRI uses multiple sliced images for a particular region in medical anatomical landmark, the approach of sample wise normalisation helps better class distinctive determination with focus of feature selective ability and this assignment use with permutation method where there is selective feature model performance is checked, so using a sample wise normalisation is best approach.