TreeMoCo: Contrastive Neuron Morphology Representation Learning

Authors: Hanbo Chen, Jiawei Yang, Daniel Iascone, Lijuan Liu, Lei He, Hanchuan Peng, Jianhua Yao

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We test Tree Mo Co on 2403 high-quality 3D neuron reconstructions of mouse brains from three different public resources. Our results show that Tree Mo Co is effective in both classifying major brain cell-types and identifying sub-types.
Researcher Affiliation Collaboration Hanbo Chen , Tencent AI Lab hanbochen@tencent.com Jiawei Yang , University of California, Los Angeles jiawei118@ucla.edu Daniel Maxim Iascone University of Pennsylvania, Philadelphia daniel.iascone@pennmedicine.upenn.edu Lijuan Liu Southeast University, Nanjing juan-liu@seu.edu.cn Lei He University of California, Los Angeles lhe@ee.ucla.edu Hanchuan Peng Southeast University, Nanjing h@braintell.org Jianhua Yao Tencent AI Lab jianhuayao@tencent.com
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https: //github.com/Tencent AILab Healthcare/Neuron Representation.
Open Datasets Yes We download data from 3 public resources: (1) BICCN f MOST data from Brain Image Library (BIL, https://download.brainimagelibrary.org/biccn/zeng/luo/f MOST/) [28], (2) Janelia Mouse Light (JML, http://mouselight.janelia.org/) [44], and (3) Allen Cell Types (ACT, https://celltypes.brain-map.org/) [1].
Dataset Splits No The paper specifies a train/test split (80% training, 20% testing) but does not explicitly mention a separate validation set or split percentage for validation data.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU models, CPU types, memory) used to run the experiments within the provided text.
Software Dependencies No The paper refers to implementation details and settings in Appendix B.1 and B.2, but the provided text does not explicitly list specific software dependencies with version numbers.
Experiment Setup Yes All models share the same 29-d input features and are trained for 100 epochs. We set the number of neighbors to 20 for BIL-6 and ACT datasets and 5 for the JML-4 dataset since a minority class might only have 12 samples in total in JML-4.