MonoBox: Tightness-Free Box-Supervised Polyp Segmentation Using Monotonicity Constraint
Authors: Qiang Hu, Zhenyu Yi, Ying Zhou, Fan Huang, Mei Liu, Qiang Li, Zhiwei Wang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on both public synthetic and in-house real noisy datasets demonstrate that Mono Box exceeds other anti-noise state-of-the-arts by improving Dice by at least 5.5% and 3.3%, respectively. |
| Researcher Affiliation | Collaboration | 1Wuhan National Laboratory for Optoelectronics, Huazhong University of Science and Technology 2 School of Engineering Sciences, Huazhong University of Science and Technology 3 Wuhan United Imaging Healthcare Surgical Technology Co., Ltd. 4 Tongji Medical College, Huazhong University of Science and Technology |
| Pseudocode | No | The paper describes the method using equations and textual descriptions of processes, but no explicit pseudocode block or algorithm is presented. |
| Open Source Code | Yes | Code https://github.com/Huster-Hq/Mono Box |
| Open Datasets | Yes | Public Synthetic Noisy Dataset. We select five public polyp datasets used in previous work (Wang et al. 2023): Clinic DB (Bernal et al. 2015), Kvasir-SEG (Jha et al. 2020), Colon DB (Tajbakhsh, Gurudu, and Liang 2015), Endo Scene (V azquez et al. 2017), and ETIS (Silva et al. 2014). |
| Dataset Splits | Yes | We set 550 samples in Clinic DB and 900 samples in Kvasir as the train-set, and the remaining samples from these two datasets and all samples from the other three datasets as the test-set. The in-house dataset cosists 18,656 colonoscopy images of polyp. The dataset is from a local hospital, which is split to the train-set (17,350 images) and test-set (1,306 images). |
| Hardware Specification | Yes | Mono Box is implemented using Py Torch (Paszke et al. 2019) and trained using a single NVIDIA Ge Force RTX 3090 GPU with 24GB memory. |
| Software Dependencies | No | The paper mentions "Py Torch" but does not specify a version number for this or any other key software dependency. |
| Experiment Setup | Yes | We use Adam W (Loshchilov and Hutter 2017) as the optimizer, and set both the learning rate and weight decay to 0.0001. We resize the input image into 352 352 and set the batch size to 16. We train the models for 50 epochs in total and evoke the label correction every 10 epochs, i.e., t = 10. The unconfident scale λ in Eq. 3 and the Io U threshold τ in label correction are set to 0.2 and 0.7, respectively. |