Stabilized Medical Image Attacks

Authors: Gege Qi, Lijun GONG, Yibing Song, Kai Ma, Yefeng Zheng

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on several medical image analysis benchmarks including the recent COVID-19 dataset show the stability of the proposed method.
Researcher Affiliation Industry Gege Qi1, Lijun Gong1 , Yibing Song2 , Kai Ma1, Yefeng Zheng1 1 Tencent Jarvis Lab, Shenzhen, China 2 Tencent AI Lab, Shenzhen, China
Pseudocode Yes We provide more results including using different Gaussian kernels W and show the pseudo code in the supplementary files.
Open Source Code Yes The code is available at https://github.com/ imogenqi/SMA
Open Datasets Yes We use two datasets for diabetic retinopathy grading. One is the APTOS-2019 (APT, 2019) dataset with 3,662 fundus images. The other is a large-scale Kaggle-DR (Kag, 2015) dataset where we randomly select 11,000 fundus images from its original training set. Both APTOS-2019 and Kaggle DR contains five defined categories. For artefact detection we use EAD-2019 (EAD, 2019) dataset with 2,500 images collected from endoscopic video frames... For lung segmentation, we use the COVID-19 dataset (COV, 2019).
Dataset Splits No The paper mentions using a 'training set' for Kaggle-DR but does not specify exact train/validation/test splits (e.g., percentages or sample counts) or any cross-validation setup for any of the datasets.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory used for running the experiments.
Software Dependencies No The paper mentions deep learning architectures like 'Res Net-50', 'graph convolutional network', 'multi-scale booster framework', and 'U-Net', but it does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes During perturbation generation, we set ϵ as 0.05, 0.01, 5 × 10−5 for lung segmentation, artefact detection and diabetic retinopathy grading respectively. We stop at the 10-th iteration for all the attack methods. The α in Eq. 4 is set as 1.