Learning Generalized Medical Image Segmentation from Decoupled Feature Queries
Authors: Qi Bi, Jingjun Yi, Hao Zheng, Wei Ji, Yawen Huang, Yuexiang Li, Yefeng Zheng
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show its state-of-the-art performance, notably outperforming the runner-up by 1.31% and 1.98% with DSC metric on generalized fundus and prostate benchmarks, respectively. |
| Researcher Affiliation | Collaboration | 1Jarvis Research Center, Tencent You Tu Lab, Shen Zhen, China 2School of Remote Sensing and Information Engineering, Wuhan University, Wuhan, China 3Department of Electrical and Computer Engineering, University of Alberta, Edmonton, Canada 4Medical AI Re Search (MARS) Group, Guangxi Medical University, Nanning, China |
| Pseudocode | No | The paper describes its method through textual descriptions and mathematical equations but does not provide any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Source code is available at https: //github.com/Bi Qi WHU/DFQ. |
| Open Datasets | Yes | DG Fundus benchmark (Wang et al. 2020) consists of four optic cup/disc segmentation datasets, namely, Drishti GS (Sivaswamy et al. 2015), RIM-ONE-r3 (Fumero et al. 2011), REFUGE (train) (Orlando et al. 2020), and REFUGE (val) (Orlando et al. 2020), which we denote as Domain-1, Domain-2, Domain-3 and Domain-4, respectively. DG Prostate benchmark (Liu, Dou, and Heng 2020) consists of 116 T2-weighted MRI cases from six domains, which we denote from Domain-1 to Domain-6. |
| Dataset Splits | Yes | DG Fundus benchmark (Wang et al. 2020) consists of four optic cup/disc segmentation datasets, namely, Drishti GS (Sivaswamy et al. 2015), RIM-ONE-r3 (Fumero et al. 2011), REFUGE (train) (Orlando et al. 2020), and REFUGE (val) (Orlando et al. 2020), which we denote as Domain-1, Domain-2, Domain-3 and Domain-4, respectively. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments, such as GPU models, CPU types, or memory amounts. |
| Software Dependencies | No | The paper mentions using 'Mix Transformer (Mi T-B3)' and 'Transformer segmentation models' but does not specify software versions for programming languages, libraries, or frameworks used in the implementation (e.g., Python, PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | Mix Transformer (Mi T-B3) (Xie et al. 2021) is used as the backbone. For the final MLP before the segmentation head, the embedding dimension is set 768. Following prior work (Zhou, Qi, and Shi 2022), the model was trained 400 epochs with an initial learning rate 5 10 4 on the Fundus benchmark, and 200 epochs with an initial learning rate 3 10 4 on the Prostate benchmark. The data pre-processing strictly follows the prior works (Wang et al. 2020; Zhou, Qi, and Shi 2022), where the fundus images were firstly centered cropped into a size of 800 800 pixels. Both prostate images and the cropped fundus images were resized into 256 256 pixels as input. |