Cell Graph Transformer for Nuclei Classification

Authors: Wei Lou, Guanbin Li, Xiang Wan, Haofeng Li

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results suggest that the proposed cell graph transformer with topology-aware pretraining significantly improves the nuclei classification results, and achieves the stateof-the-art performance. Code and models are available at https://github.com/lhaof/CGT
Researcher Affiliation Academia 1Shenzhen Research Institute of Big Data, Shenzhen, China 2The Chinese University of Hong Kong, Shenzhen, China 3School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China 4Guang Dong Province Key Laboratory of Information Security Technology
Pseudocode No The paper does not contain any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Code and models are available at https://github.com/lhaof/CGT
Open Datasets Yes We utilize four nuclei classification datasets: Pan Nuke (Gamper et al. 2020), Lizard (Graham et al. 2021), Nu CLS (Amgad et al. 2022), and BRCA-M2C (Abousamra et al. 2021).
Dataset Splits Yes The data split and more details are in the supplementary material.
Hardware Specification Yes The pretraining strategy and the training of CGT are run for 150 and 50 epochs, respectively, with the Adam optimizer in an NVIDIA A-100 GPU.
Software Dependencies No The implementation is based on Py Torch (Paszke et al. 2017) and Py Torch Geometric library (Fey and Lenssen 2019). Specific version numbers for these libraries are not provided.
Experiment Setup Yes For the proposed CGT, the encoder and decoder of the feature extractor have four layers and three layers, respectively. The CGT encoder contains four transformer layers. For the pretraining strategy, the GCN is built of two GENConv (Li et al. 2020) layers. Our results are reported as the average result of training with three different random seeds. The dimensions of type markers and link markers are 64 and 16. The number of edges of each node is 4. The pretraining strategy and the training of CGT are run for 150 and 50 epochs, respectively, with the Adam optimizer in an NVIDIA A-100 GPU. The initial learning rates for pretraining and training are 10 4 and 10 5, respectively.