Efficient Folded Attention for Medical Image Reconstruction and Segmentation
Authors: Hang Zhang, Jinwei Zhang, Rongguang Wang, Qihao Zhang, Pascal Spincemaille, Thanh D. Nguyen, Yi Wang10868-10876
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the superiority of our method on two challenging tasks for 3D MIR and MIS, which are quantitative susceptibility mapping and multiple sclerosis lesion segmentation. [...] Extensive experimental results from both a semantic segmentation task and a functional image mapping task on 3D medical images show the effectiveness and the efficiency of our method. |
| Researcher Affiliation | Academia | 1 Cornell University, Ithaca NY, USA 2 Weill Cornell Medical College, New York NY, USA 3 University of Pennsylvania, Philadelphia PA, USA |
| Pseudocode | No | The paper describes the proposed method using mathematical equations and diagrams (Figure 3), but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our implementation is made publicly available 1. 1https://github.com/tinymilky/FANet |
| Open Datasets | No | For MS lesion segmentation, 'We use a dataset with 30 MR images acquired from a 3.0 T GE scanner.' For QSM, 'To acquire and reconstruct COSMOS data, 6 healthy subjects were recruited to do MRI scan with 5 brain orientations using a 3.0T GE scanner.' No concrete access information for a publicly available or open dataset is provided. |
| Dataset Splits | Yes | We perform five random splits on the dataset, where each split contains 15, 5, and 10 subjects for training, validation, and testing. [...] We perform six splits on the dataset, where each split contains 4, 1, and 1 subjects(s) for training, validation, and testing, and each subject contains 5 volumes. |
| Hardware Specification | Yes | All models in the experiments are trained in a machine with a Titan Xp GPU. |
| Software Dependencies | No | The paper states 'We use Py Torch (Paszke et al. 2019) for all of our implementations.' but does not provide a specific version number for PyTorch or any other software dependency. |
| Experiment Setup | Yes | For MS lesion segmentation: 'We perform random crop with fixed cropping size (128 160 32), and use elastic deformation, intensity shifting for data augmentation. We adopt the sum of weighted cross entropy and soft dice (Dice 1945) as our loss function. Adam (Kingma and Ba 2014) with the initial learning rate of 1e 3 and a multi-step learning rate scheduler with milestones at 50%, 70% and 90% of the total epochs are used for optimal convergence. A batch size of four is used for training, and training would stop after 120 epochs.' For QSM: 'During training, we cropped each volume into 3D patches in size (64 64 32) and use in-plane rotation of 15 for data augmentation. Loss function from QSMnet (Yoon et al. 2018) is adopted. Adam (Kingma and Ba 2014) optimizer is used for training with the same hyperparameters as MS lesion segmentation experiment. Training is performed with a batch size of 16 and training would stop after 60 epochs.' |