BGDB: Bernoulli-Gaussian Decision Block with Improved Denoising Diffusion Probabilistic Models
Authors: Chengkun Sun, Jinqian Pan, Russell Stevens Terry, Jiang Bian, Jie Xu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed method across various imaging tasks, including both classification and segmentation. ... Experimental Results As illustrated in Table 1, on both Cityscapes and Pascal VOC, all models experienced moderate improvements. ... Ablation Experiments To thoroughly understand the impact of different loss functions on the performance of our model, we conducted ablation experiments using the U-Net model with various combinations of loss functions. |
| Researcher Affiliation | Academia | 1Department of Health Outcomes and Biomedical Informatics, University of Florida, Gainesville, FL 32611, USA 2Department of Urology, University of Florida, Gainesville, FL 32611, USA EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology using mathematical equations and prose, and includes a 'Figure 1: Workflow for performing segmentation tasks', but does not contain any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code https://github.com/sunck1/BGDB |
| Open Datasets | Yes | Urban and General Scene Segmentation Datasets We utilized Cityscapes (Cordts et al. 2016) and PASCAL Visual Object Classes (VOC) Challenge (Pascal VOC) (Everingham et al. 2010) datasets for this task. ... Skin Lesion Segmentation Datasets We used the International Skin Imaging Collaboration (ISIC) dataset (Tschandl, Rosendahl, and Kittler 2018; Codella et al. 2019) for skin lesion segmentation. |
| Dataset Splits | Yes | We utilized the training, validation, and test datasets provided by the ISIC 2018 challenge. These datasets were combined and then randomly split into training and testing sets in a 5:2 ratio (2,600 images for training and 1,094 for testing). We performed 5-fold cross-validation, selecting the optimal model from each fold s validation set. |
| Hardware Specification | Yes | In this study, all models were trained on an NVIDIA A100 GPU with 80 GB of memory. |
| Software Dependencies | No | We evaluated the Bernoulli Gaussian decision block across several classical and state-of-the-art 2D medical segmentation models using the ISIC dataset. These models include U-Net (Ronneberger, Fischer, and Brox 2015), Attention U-Net (Oktay et al. 2018), U-Net++ (Zhou et al. 2019), FCN (Liu et al. 2018), Res UNet (Diakogiannis et al. 2020), and UNETR (Hatamizadeh et al. 2022), all implemented using the MONAI framework (Cardoso et al. 2022). The paper mentions the MONAI framework but does not specify a version number. |
| Experiment Setup | Yes | Our training regimen consisted of 30,000 iterations, with each batch comprising 16 samples. All input images were uniformly cropped to dimensions of 256 × 256. We employed the cross-entropy loss function, coupled with a learning rate of 0.01 and a weight decay of 1e-4. Stochastic Gradient Descent (SGD) (Robbins and Monro 1951) was used as the optimizer throughout the training process... The hyperparameters were set as follows: λ1 to 1 × 10−3, and both λ2 and λ3 to 1. |