Transfusion: Understanding Transfer Learning for Medical Imaging

Authors: Maithra Raghu, Chiyuan Zhang, Jon Kleinberg, Samy Bengio

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental A performance evaluation on two large scale medical imaging tasks shows that surprisingly, transfer offers little benefit to performance, and simple, lightweight models can perform comparably to IMAGENET architectures.
Researcher Affiliation Collaboration Maithra Raghu Cornell University and Google Brain chiyuan@google.com Chiyuan Zhang Google Brain chiyuan@google.com Jon Kleinberg Cornell University kleinber@cs.cornell.edu Samy Bengio Google Brain bengio@google.com
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement or a direct link for the open-sourcing of the code for its methodology. It mentions using 'Tensorflow Slim' for pretrained weights, but this refers to a third-party library, not their own implementation code.
Open Datasets Yes Our primary dataset, the RETINA data, consists of retinal fundus photographs [9]... We also study a second medical imaging dataset, CHEXPERT [14], which consists of chest x-ray images...
Dataset Splits No The paper mentions training on various datasets, including using "5000 datapoints on the RETINA dataset" for the small data regime, and states that models are evaluated on medical tasks. However, it does not explicitly specify the percentages or absolute counts for train, validation, or test splits. It refers to "three repetitions of the different models" but this indicates training runs, not data partitioning.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments, such as exact GPU or CPU models, processor types, or memory specifications.
Software Dependencies No The paper mentions using "Tensorflow Slim" [26] for pretrained Inceptionv3 weights, but it does not specify the version numbers for TensorFlow or any other software libraries or dependencies essential for reproducing the experiments.
Experiment Setup No The paper describes the general experimental procedure, such as training models from random initialization or using transfer learning, and the types of architectures used (ResNet50, Inception-v3, CBR family). However, it does not provide specific experimental setup details, such as hyperparameter values (e.g., learning rates, batch sizes, optimizers, or number of epochs), that are necessary for reproducing the results.