Datasets
Standard Dataset
ROBUSTNESS BENCHMARK DATASETS FOR SEMANTIC SEGMENTATION OF FLUORESCENCE IMAGES UPDATED
- Citation Author(s):
- Submitted by:
- Liqun Zhong
- Last updated:
- Sat, 06/22/2024 - 00:57
- DOI:
- 10.21227/1jk9-nv64
- License:
117 Views- Categories:
- Keywords:
Abstract
We have developed three datasets, referred to as ER-C, Mito-C and Nucleus-C, respectively, for benchmarking robustness of DNN models against corruptions and adversarial attacks in semantic segmentation of fluorescence microscopy images. Degraded images in these three datasets are synthesized from raw images along with their manually annotated segmentation labels in the ER, Mito, and Nucleus datasets [1,2], respectively. They are synthesized with controlled corruptions and adversarial attacks. The initial version of this dataset was released in 2022 [3] (https://ieee-dataport.org/documents/robustness-benchmark-datasets-semantic-segmentation-fluorescence-images). In this updated version, the pure Gaussian noise used in the initial version has been changed to mixed Poisson-Gaussian noise.
[1] Y. Luo, Y. Guo, W. Li, G. Liu, and G. Yang. Fluorescence Microscopy Image Datasets for Deep Learning Segmentation of Intracellular Orgenelle Networks [Online] Available: https://dx.doi.org/10.21227/t2he-zn97.
[2] J. C. Caicedo et al., "Nucleus segmentation across imaging experiments: the 2018 Data Science Bowl," Nature Methods, vol. 16, no. 12, pp. 1247-1253, 2019.
[3] Liqun Zhong, Ge Yang, Lingrui Li, June 26, 2022, "Robustness Benchmark Datasets for Semantic Segmentation of Fluorescence Images", IEEE Dataport, doi: https://dx.doi.org/10.21227/kxay-5y38.
The overall workflow of our image synthesis consists of three steps (refer to [4] for more details). First, segmentation labels are used as binary masks to guide synthesis of images using a generative adversarial network (GAN), which are trained to learn the mapping from the masks to their corresponding fluorescence microscopy images. The masks are used as the ground truth for the final output images. Second, denoising is performed on the synthesized images to remove their background noise using the method in [5]. This step is important because it enables precise control of signal-to-noise ratios (SNRs) in the next step. Third, different corruptions and adversarial attacks are applied to the denoised synthetic images to generate the final output images for benchmarking robustness of DNN models. Specifically, we simulate six SNR levels (SNR=1,2,3,4,5,8), five space-invariant blurring (SIB) levels and five space-variant blurring (SVB) levels.
[4] Liqun, Zhong; Li, Lingrui; Yang, Ge (2022): Characterizing Robustness of Deep Neural Networks in Semantic Segmentation of Fluorescence Microscopy Images. TechRxiv. Preprint. https://doi.org/10.36227/techrxiv.20188742.vl
[5] L. Zhong, G. Liu, and G. Yang, "Blind Denoising of Fluorescence Microscopy Images Using GAN-Based Global Noise Modeling," in International Symposium on Biomedical Imaging (ISBI), 2021: IEEE, pp. 863-867.
