Contents

Neural-network based picking

Neural-network based picking

nextPYP implements semi-supervised particle picking using neural-networks both for 2D micrographs and 3D tomograms

Step 1: Pick particles for training

  • Click inside the Pre-processing block and go to the Micrographs tab

  • Create a new list by entering a name and clicking New

  • Select particles in the current micrograph by clicking on their centers

  • Navigate to other micrographs in the dataset and select additional particles as needed

Create new filter

  • Click inside the Pre-processing block, go to the Tilt-series tab, and select the Reconstruction group

  • Create a new list by entering a name and clicking New

  • Select particles in the current tomogram by clicking on their centers. Use the slider below the image to scroll through the tomogram

  • Navigate to other tomograms in the dataset and select additional positions as needed

Create new filter

Note

  • Particles can be deleted by right-clicking on the markers

  • Particle positions are saved automatically every time a particle is added or deleted

  • The total number of particles in a dataset is displayed on the top-left corner of the page

Tip

The size of the markers can be controlled by changing the Detection radius in the Particle detection tab. The block must be re-run for this change to take effect

Step 2: Train the neural-network model

  • Open the settings of the Pre-processing block, go to the Particle detection tab and select nn-train as the Detection method

  • Choose the list of manually selected positions from the Select list for training dropdown menu at the top of the form

Create new filter

  • Go to the Training/Evaluation tab and set the desired parameters for training

Create new filter

  • Click Save, then Run to train the model

Note

Since training is run using the GPU, a GPU partition must be configured in the nextPYP instance

Step 3: Run inference using the trained model

  • Go to the Particle detection tab in the Pre-processing block and select nn-eval as the Detection method

  • Go to the Training/Evaluation tab and select the location of the trained model obtained in the previous step (train/YYYYMMDD_HHMMSS/*.training for 2D, and train/YYYYMMDD_HHMMSS/*.pth for 3D)

  • Click Save, then Run to evaluate the model on all the micrographs or tomograms

  • Inspect the results using the Micrographs tab (2D) or the Reconstruction group in the Tilt-series tab (3D)

Tip

Since the quality of the picking may depend on the size of the training set, challenging datasets may require the use of more particles for training

See also

previous

Particle picking

next

Reference