Finding the best preprocessing

[Luugaaa]

Context

In order to improve the results, we want to find the best preprocessing for the magnitude and the phase. A good preprocessing should facilitate the segmentation for the model from the very beginning. To evaluate the best preprocessing, we conduct experiments based on the default implementation over 25 epochs.

Preprocessing techniques

In the default implementation, there is no heavy preprocessing meant to improve the quality of the data given to the model. Here, for now, we investigate the impact of both the use CLAHE on the magnitude channel and/or the use of a linear contrast enhancement on the phase. We evaluate the impact on the validation metrics. Naturally, the preprocessing is applied on both train and val data and the training is deterministic.

Note for later : for now we implement the preprocessing as an additional step done after augmentation. For a better consistency, it could/should be integrated in the get_training_transforms and the get_val_transforms.

[Luugaaa]

Note : for the following experiments, the mag preprocessing is

CLAHE(mag, clip_limit=0.01)

and the phase preprocessing is :

p2, p98 = np.percentile(unwrapped_phase, (30, 75))
rescale_intensity(unwrapped_phase, in_range=(p2, p98), out_range=(0, 1))

First, we applied both the phase preprocessing and the magnitude preprocessing (red curve, Exp_2_default_prepro_mag_phase) and compared the results with the same implementation with no preprocessing applied (green curve, Exp_2_default_prepro_mag_phase).

At first the training with both prepro has worse results than that default training, but the gap seems to be closed toward the end. This is promising, we might want to test over more epochs.

On each class we have the following results :

• class 0 : bg
• class 1 : wm
• class 2 : gm
• class 3 : lesion_wm
• class 4 : lesion_gm

Most difference is found in the segmentation of GM and WM. For GM, the preprocessing seems to greatly increase the results (+0.1 on dice), for the WM it's the other way around, we lose around 0.1 with the preprocessing.

When removing the phase prepro, leaving only the magnitude prepro, there is a significant decrease in the mean dice (see below).

For each class, both the dual preprocessing and the mag only preprocessing seems to yield similar results (see below), except for the GM segmentation. This seems to indicate the phase preprocessing is behind the improvement for the GM segmentation (which actually not really surprising). On the other hand, the WM segmentation stays low, this could indicate the mag preprocessing is the source of the dice decrease for this class.

When preprocessing only the phase (orange curve, Exp_4_default_prepro_phase), the WM segmentation dice goes back to what is was without prepro, actually with a slight increase. Surprisingly, for the GM segmentation we don't observe the dice increase we had with both preprocessing.

The mean val dice is satisfying without showing any significant increase.

As a partial conclusion, this mag preprocessing (clahe with the 0.01 clip_limit) doesn't seem relevant as there is no improvement what so ever on any metric, but actually a decrease in some. The phase preprocessing seems more relevant but should be approach with care, further investigation is needed.

For the record, here is a visual demonstration of the applied preprocessing.
Original images :

Preprocessed images :

The preprocessing of the noisy phase image created strong artifacts in the background while the mag preprocessing introduced a slight noise.

Next step :
Without necessarily changing the types of processing used, we could try to clean the images of the noise in the background to limit the apparition of artifacts.