Detnet: Deep Neural Network For Particle Detection In Fluorescence Microscopy Images

What is this post about?

• Notes from reading the 2019 paper
  • Detnet: Deep Neural Network For Particle Detection In Fluorescence Microscopy Images
  • by T. Wollmann, C. Ritter, J. N. Dohrke, J.-Y. Lee, R. Bartenschlager, K. Rohr
  • https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8759234

Short take

• Hourglass-shaped network for detection of spots/foci/particles in cells
• Based on Deconvolution network with the following changes
  • Reduced number of feature maps
  • Reduces size of receptive field (two instead of five poolings)
  • No long-range skip connections (detailed boundary information is irrelevant)
  • Use residual blocks instead of convolution (to reduce issue of vanishing gradient)
  • Use instance- instead of batch-normalisation (because little training data)
  • Use bilinear upsampling instead of transposed convolutions (expanding path)
  • 17k parameters instead of U-net’s 1.9M
  • Use soft Dice loss (performs implicit class balancing and penalizes easy samples)
  • Use early stopping
  • Availability
    • No information given, no links, not GitHub names
    • Contact with first-author points to company in Berlin
      • https://www.merantix.com/

Miscellaneous

• Details
  • AMSGrad optimizer
  • in-house hyperparameter optimization framework HyperHyper
• Quotes
  • domain adapted Deconvolution Network
  • can be trained with only a few ground truth annotations
  • benchmarked using data from the ISBI Particle Tracking Challenge
  • including particles with different shapes (round and elongated)
  • when using a standard Cross-Entropy loss, training was not successful due to the heavy class imbalance
  • the F1 score can be significantly improved by optimizing the shift a ∈ R of the sigmoid function of the neural network
  • The assignment between particle detections and ground truth was determined by the Munkres algorithm with a maximal gating distance of 5 pixels.