Exploring ML in Cryo-EM

This week I worked on finalizing the regression model, testing out different combinations of bandpass filters, models, and noise. I trained on the fixed apoferritin dataset. This graph is of one of the best models trained so far:

Since the labels are the sum of 4 quality metrics, I compared the graph of the predictions with each individual label:

I also ran a modified SRGAN to get some evaluation results and look at the feasibility of using it for denoising:

Ground truth and generated image

Next week, I plan on incorporating the power spectrum into the regression model, evaluating the SRGAN, and summarize our results.

Started training SRGAN on the t20s dataset with correct projections. Also added an additional loss component that finds the MSE between the feature maps of a pretrained VGG model fed the generated data, and a pretrained VGG model that is fed the upsampled low pass of the original image patch. This was to ensure that the generated image preserved some of the features and details of the original image.

Also looked at how each individual label correlated with the predictions:

Added gaussian noise to input images as a form of data augmentation in hopes of reducing overfitting. The problem is that adding noise would theoretically reduce the quality score, but I gave it a shot anyway

• Gaussian noise with .05 std (images are normalized to a range of 0-1): eval loss increased slightly
• Gaussian noise with .02 std: eval loss stayed roughly the same

Created a class to load a model and its weights to use for prediction. Gave model/weights trained on single images of apoferritin dataset to Xiaochen for his alignment model. However, the aligned reconstruction worse than the original. There may have been a problem with the apoferritin dataset, so I retrained the model on the ts20 dataset. However, the model performs much worse:

Tested various regression models:

• 3 filtered input “residual network”
• Single image network
  • lowpass image
  • original image
• Projection model
  • lowpass image + projection
  • original image + projection
• Combined
  • 3 filtered images + projection

Lowpass image + projection worked the best with a loss of 0.05, however, all the other models worked well too, with a loss ranging from .06 to .1

Tested various regression models:

• 3 filtered input “residual network”
• Single image network
  • lowpass image
  • original image
• Projection model
  • lowpass image + projection
  • original image + projection
• Combined
  • 3 filtered images + projection

Lowpass image + projection worked the best with a loss of 0.05, however, all the other models worked well too, with a loss ranging from .06 to .1

Tested various regression models:

• 3 filtered input “residual network”
• Single image network
  • lowpass image
  • original image
• Projection model
  • lowpass image + projection
  • original image + projection
• Combined
  • 3 filtered images + projection

Lowpass image + projection worked the best with a loss of 0.05, however, all the other models worked well too, with a loss ranging from .06 to .1