Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Scale-Aware Contrastive Reverse Distillation for Unsupervised Medical Anomaly Detection
Authors: Chunlei Li, Yilei Shi, Jingliang Hu, Xiaoxiang Zhu, Lichao Mou
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on benchmark datasets demonstrate stateof-the-art performance, validating the efficacy of the proposed method. Code is available at url https://github.com/Med AITech/SCRD4AD. |
| Researcher Affiliation | Collaboration | Chunlei Li Med AI Technology (Wuxi) Co. Ltd. EMAIL Yilei Shi Med AI Technology (Wuxi) Co. Ltd. EMAIL Jingliang Hu Med AI Technology (Wuxi) Co. Ltd. EMAIL Xiao Xiang Zhu Technical University of Munich EMAIL Lichao Mou Med AI Technology (Wuxi) Co. Ltd. EMAIL |
| Pseudocode | Yes | Algorithm. We present detailed procedures for synthesizing abnormal images and training our model in Algorithm 1 and Algorithm 2, respectively. Algorithm 1 Synthesize abnormal images Algorithm 2 Pseudo-code of our approach in one epoch training |
| Open Source Code | Yes | Code is available at url https://github.com/Med AITech/SCRD4AD. |
| Open Datasets | Yes | We evaluate our proposed method on three widely-used medical imaging datasets: the RSNA Pneumonia Detection Challenge dataset*, the Brain Tumor MRI dataset , and the ISIC 2018 dataset . *https://www.kaggle.com/c/rsna-pneumonia-detection-challenge https://www.kaggle.com/datasets/masoudnickparvar/ brain-tumor-mri-dataset https://challenge.isic-archive.com/data/#2018 https://www.kaggle.com/datasets/ahmedhamada0/brain-tumor-detection |
| Dataset Splits | Yes | RSNA: This chest X-ray dataset comprises 8,851 normal and 6,012 lung opacity images. Following Cai et al. (2022), we utilize 3,851 normal images for training and a balanced test set of 1,000 normal and 1,000 abnormal images. Brain Tumor: ...our experimental setup includes 1,000 normal slices for training and a test set of 600 normal and 600 abnormal slices (equally split between glioma and meningioma). ISIC: ...employs 6,705 normal images from the official training set for model training. Our test set comprises 909 normal images and 603 abnormal images (distributed across the remaining six categories) from the official test set. |
| Hardware Specification | Yes | We conduct all experiments using Py Torch on a single NVIDIA RTX 3090Ti GPU. |
| Software Dependencies | No | The paper mentions 'Py Torch' but does not specify a version number. No other specific software dependencies with version numbers are provided. |
| Experiment Setup | Yes | To train our model, we employ the Adam optimizer (Kingma & Ba, 2015) with β = (0.5, 0.999) and a learning rate of 1e-3. We train for 4,000 iterations with a batch size of 16. Our decoder mirrors the encoder, identical to that used in RD4AD (Deng & Li, 2022). For competing methods, we utilize their publicly available codes and adhere to their default training configurations. In our approach, for each training image x, we first sample noise size and position from uniform distributions. Then, simplex noise (Perlin, 2002) is generated with six octaves and a persistence of γ = 0.6. Finally, the generated noise is added to the image, scaled by a factor of λ = 0.2. This process synthesizes an abnormal image x , introducing structured perturbations that mimic potential anomalies. We resize all images to 256 256 pixels and apply no data augmentation during training. |