google/ffn

Flood-Filling Networks for instance segmentation in 3d volumes.

What it solves

Flood-Filling Networks (FFNs) are designed for instance segmentation of complex, large shapes within volumetric datasets, specifically targeting brain tissue volume EM (electron microscopy) datasets.

How it works

FFNs use a specialized neural network architecture to iteratively segment objects. The project provides tools to prepare training data by computing "partitions" (quantized fractions of labeled voxels within a specific radius) and generating coordinate files for sampling. Training is performed via a train.py script using TensorFlow, and inference can be run either as a batch process to segment entire volumes or interactively via a seed-based approach to grow a segmentation mask.

Who it’s for

Researchers and data scientists working with 3D volumetric imaging, specifically those focused on biological imaging and electron microscopy of brain tissue.

Highlights

  • Volumetric Segmentation: Specifically optimized for large and complex 3D shapes.
  • Data Preparation Pipeline: Includes scripts to compute partitions and build coordinate files for efficient training.
  • Training and Inference: Supports both non-interactive batch inference and interactive seed-based segmentation.
  • GPU Acceleration: Optimized for use with Tesla P100 GPUs and similar hardware for computationally expensive volumetric processing.

Related

  • Project
    flexflow/flexflow-train

    A deep learning framework that accelerates distributed DNN training by automatically searching for the most efficient parallelization strategies.

  • Project
    Xilinx/finn

    An experimental framework from AMD Research that enables fast, scalable inference of quantized neural networks on FPGAs by generating customized dataflow architectures.

  • Project
    AIStream-Peelout/flow-forecast

    An open-source deep learning framework for time series forecasting, classification, and anomaly detection featuring a wide array of state-of-the-art models.

  • Project
    Deep-MI/FastSurfer

    A fast, deep-learning based neuroimaging pipeline for volumetric and surface-based thickness analysis of brain MRI images, serving as a high-speed alternative to FreeSurfer.

  • Project
    google-research/morph-net

    A framework for learning deep network structure during training to reduce FLOPs, model size, and latency through resource-constrained regularization.