Learning Low-Rank Feature for Thorax Disease Classification
Authors: Yancheng Wang, Rajeev Goel, Utkarsh Nath, Alvin Silva, Teresa Wu, Yingzhen Yang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We study thorax disease classification in this paper. ... In the empirical study, using a neural network such as a Vi T or a CNN pre-trained on unlabeled chest X-rays by Masked Autoencoders (MAE), our novel LRFL method is applied on the pre-trained neural network and demonstrates better classification results in terms of both multi-class area under the receiver operating curve (m AUC) and classification accuracy than the current state-of-the-art. |
| Researcher Affiliation | Academia | Yancheng Wang1 Rajeev Goel1 Utkarsh Nath1 Alvin C. Silva2 Teresa Wu1 Yingzhen Yang1 1 School of Computing and Augmented Intelligence, Arizona State University {ywan1053, rgoel15, unath, teresa.wu, yingzhen.yang}@asu.edu 2 Mayo Clinic Arizona silva.alvin@mayo.edu |
| Pseudocode | Yes | Algorithm 1 describes the training algorithm for the neural network trained with LRFL, which uses the standard SGD to optimize the loss function LLRFL(W), as step three of our pipeline in Section 3.1. |
| Open Source Code | Yes | The code of LRFL is available at https://github.com/ Statistical-Deep-Learning/LRFL. |
| Open Datasets | Yes | We fine-tune these pre-trained networks with low-rank regularization on three public X-ray datasets: (1) NIH Chest X-ray14 [10], (2) Stanford Che Xpert [12], and (3) COVIDx [11]. |
| Dataset Splits | Yes | The value of γ is selected from {0.01, 0.02, 0.03, 0.04, 0.05, 0.1, 0.15, 0.2}. The value of η is selected from 5 10 4, 1 10 3, 2.5 10 3, 5 10 3, 1 10 2 . The value of µ is selected form 5 10 4, 2.5 10 4, 1 10 4, 5 10 5, 2.5 10 5, 1 10 5 . To determine the optimal values of the parameters η, γ, and µ, we employ a sequential greedy search strategy. We first fix η and µ and find the optimal value of γ by cross-validation. Subsequently, using this optimized γ, we proceed to search for the optimal η while keeping µ constant. Finally, with optimal γ and η, we search for the optimal µ by cross-validation. The optimal values of η, γ, and µ selected by cross-validation are shown in Table 8 in Section B.3 of the appendix. |
| Hardware Specification | Yes | The evaluation of LRFL models and baseline models is performed on 4 Nvidia A100 GPUs. |
| Software Dependencies | No | The paper mentions software components like 'ADAM optimizer' and 'Masked Autoencoders (MAE)' but does not provide specific version numbers for these or other underlying libraries/frameworks (e.g., PyTorch, TensorFlow). |
| Experiment Setup | Yes | The ADAM optimizer is used with a batch size of 1024 for all datasets. Initially, we fine-tune the entire networks for 75 epochs following the settings in [2], then fine-tune with low-rank regularization for another 75 epochs. We use a cosine learning rate schedule, and the initial learning rate, which is denoted as µ, is selected by cross-validation for each model and each dataset. The default values for momentum and weight decay are set to 0.9 and 0, respectively. We use standard data augmentation techniques, including random-resize cropping, random rotation, and random horizontal flipping. |