Parallel Multi-Dimensional LSTM, With Application to Fast Biomedical Volumetric Image Segmentation

Authors: Marijn F. Stollenga, Wonmin Byeon, Marcus Liwicki, Jürgen Schmidhuber

NeurIPS 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach on two 3D biomedical image segmentation datasets: electron microscopy (EM) and MR Brain images.
Researcher Affiliation Collaboration 1Istituto Dalle Molle di Studi sull Intelligenza Artificiale (The Swiss AI Lab IDSIA) 2Scuola universitaria professionale della Svizzera italiana (SUPSI), Switzerland 3Universit a della Svizzera italiana (USI), Switzerland 4University of Kaiserslautern, Germany 5German Research Center for Artificial Intelligence (DFKI), Germany
Pseudocode No No explicit pseudocode or algorithm block was found.
Open Source Code No The paper does not provide an explicit statement or link to its open-source code.
Open Datasets Yes The EM dataset [12] is provided by the ISBI 2012 workshop on Segmentation of Neuronal Structures in EM Stacks [15]. The MR Brain images are provided by the ISBI 2015 workshop on Neonatal and Adult MR Brain Image Segmentation (ISBI NEATBrain S15) [13]. Citation [15] also includes a URL: http://tinyurl.com/d2fgh7g.
Dataset Splits No One stack is used for training, the other for testing. (for EM) and The dataset is divided into a training set with five volumes and a test set with fifteen volumes. (for MR Brain). No explicit mention of a validation split.
Hardware Specification Yes All experiments are performed on a desktop computer with an NVIDIA GTX TITAN X 12GB GPU.
Software Dependencies No The paper mentions using 'NVIDIA s cu DNN library' but does not specify its version number or any other software dependencies with versions.
Experiment Setup Yes We apply RMS-prop [17] with momentum. We use a decaying learning rate: λlr = 10 6 + 10 2 1 100 , which starts at λlr 10 2 and halves every 100 epochs asymptotically towards λlr = 10 6. Other hyper-parameters used are ϵ = 10 5, ρMSE = 0.9, and ρM = 0.9. Our networks contain three Pyra Mi D-LSTM layers. The first Pyra Mi D-LSTM layer has 16 hidden units followed by a fully-connected layer with 25 hidden units. ... The convolutional filter size for all Pyra Mi D-LSTM layers is set to 7 7. 50% dropout on fully connected layers and/or 20% on input layer.