Retrieval-Augmented Multiple Instance Learning

Authors: Yufei CUI, Ziquan Liu, Yixin Chen, Yuchen Lu, Xinyue Yu, Xue (Steve) Liu, Tei-Wei Kuo, Miguel Rodrigues, Chun Jason Xue, Antoni Chan

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evaluations conducted on WSI classification demonstrate that the proposed RAM-MIL framework achieves state-of-the-art performance in both in-domain scenarios, where the training and retrieval data are in the same domain, and more crucially, in out-of-domain scenarios, where the (unlabeled) retrieval data originates from a different domain. and 5 Experiments The proposed methodology is evaluated on whole slide image (WSI) datasets (Camelyon16 [43, 13], Camelyon17 [12], TCGA-NSCLC, CPTAC-UCEC and CPTAC-LSCC) and general MIL datasets (See results in supplemental).
Researcher Affiliation Academia 1Mila, Mc Gill University 2University College London 3City University of Hong Kong 4 Mila, Université de Montréal 5National Taiwan University 6MBZUAI
Pseudocode Yes Algorithm 1 Retrieval-Aug Mented Multiple Instance Learning (RAM-MIL) Algorithm
Open Source Code Yes Code can be found at https://github.com/ralphc1212/ram-mil.
Open Datasets Yes The proposed methodology is evaluated on whole slide image (WSI) datasets (Camelyon16 [43, 13], Camelyon17 [12], TCGA-NSCLC, CPTAC-UCEC and CPTAC-LSCC) and general MIL datasets (See results in supplemental).
Dataset Splits Yes Each result is obtained with 10-fold splits of training/validation/testing sets. and To ensure robustness and avoid the influence of outliers, each experiment is executed 10 times on randomly partitioned train/validation/test sets.
Hardware Specification No The paper does not explicitly specify the hardware used for experiments, such as specific GPU models, CPU models, or memory configurations.
Software Dependencies No The paper mentions the use of "Adam optimizer" but does not specify version numbers for any software dependencies like programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow).
Experiment Setup Yes For the backbone MIL model we use the the same parameter setup as CLAM. The model parameters are updated via the Adam optimizer with an L2 weight decay of 1e-5 and a learning rate of 2e-4. and Finally, we train a single logistic regression classifier using the merged representation. The Adam optimizer is used to update the model parameters, with a L2 weight decay of 1e-4 and a learning rate of 2e-4. The models are trained for a minimum of 40 epochs and up to a maximum of 200 epochs if the early stopping criterion is not met.