notes.txt

Hyperspectral/Material Dataset Challenges:

	- Severely inbalanced (Mostly background)
	- Object structure less important for predictions than "texture" (E.g: MUSIC_2D are petri dish samples, material has to be inferred from hyperspectral channels)

--------------------------------------------------------------------------------------------------------
Empirical Observations:

	- Using regularization such as dropout and transformations helps predictions
	- Using weighted loss is necessary due to class imbalance (MUSIC_2D: background is 90% of image)
	- MUSIC_2D: Tests on dice loss shows that its convergence is not a good indicator of performance and seems to be suffering from vanishing gradients
	- MUSIC_2D: Weighted cross entropy converges 
	- Needs low learning rates (5x10^-5)
	- U-NET architecture works greatly


--------------------------------------------------------------------------------------------------------
Research Ideas:

	- Oversampling single objects to overcome imbalance
	- Explore more augmentations
	- Patchify data (related to 1)
	- Less important: more complex loss functions for imbalance

--------------------------------------------------------------------------------------------------------

References

[Unets] https://arxiv.org/pdf/1505.04597.pdf
[Automatic Multi-organ segmentation] https://aapm.onlinelibrary.wiley.com/doi/epdf/10.1002/mp.13950
[Hyperspectral] https://arxiv.org/pdf/2303.08252v1.pdf