Medical Image Segmentation using Squeeze-and-Expansion Transformers

Authors: Shaohua Li, Xiuchao Sui, Xiangde Luo, Xinxing Xu, Yong Liu, Rick Goh

IJCAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments were performed on 2D and 3D medical image segmentation tasks: optic disc/cup segmentation in fundus images (REFUGE 20 challenge), polyp segmentation in colonoscopy images, and brain tumor segmentation in MRI scans (Bra TS 19 challenge).
Researcher Affiliation Collaboration 1Institute of High Performance Computing, A*STAR, Singapore 2University of Electronic Science and Technology of China, Chengdu, China
Pseudocode No The paper describes the architecture and components of Segtran with mathematical formulations and diagrams, but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes The source code of Segtran is released at https://github.com/askerlee/segtran.
Open Datasets Yes REFUGE20: Optic Disc/Cup Segmentation in Fundus Images. This task does segmentation of the optic disc and cup in fundus images, which are 2D images of the rear of eyes (Fig. 5). It is a subtask of the REFUGE Challenge2 [Orlando et al., 2020]... We also used two extra datasets, Drishti-GS dataset [Sivaswamy et al., 2015] and RIM-ONE v3 [Fumero et al., 2011] when training all models.
Dataset Splits Yes REFUGE20: ...1200 images were provided for training, and 400 for validation. Polyp: ...Each was randomly split into 80% training and 20% validation, and the training images were merged into one set. Bra TS19: ...335 scans were provided for training, and 125 for validation.
Hardware Specification Yes All models were trained on a 24GB Titan RTX GPU with the Adam W optimizer.
Software Dependencies No The paper mentions using "PyTorch" and a "popular library Segmentation Models.Py Torch (SMP)" for implementations, but it does not provide specific version numbers for these software components.
Experiment Setup Yes All models were trained on a 24GB Titan RTX GPU with the Adam W optimizer. The learning rate for the three transformer-based models were 0.0002, and 0.001 for the other models. On REFUGE20, all models were trained with a batch size of 4 for 10,000 iterations (27 epochs); on Polyp, the total iterations were 14,000 (31 epochs). On Bra TS19, Segtran was trained with a batch size of 4 for 8000 iterations. The training loss was the average of the pixel-wise cross-entropy loss and the dice loss. Segtran used 3 transformer layers on 2D images, and 1 layer on 3D images to save RAM. The number of modes in each transformer layer was 4.