Bi-directional Weakly Supervised Knowledge Distillation for Whole Slide Image Classification

Authors: Linhao Qu, xiaoyuan luo, Manning Wang, Zhijian Song

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on five datasets demonstrate the efficiency of WENO.
Researcher Affiliation Academia Linhao Qu Digital Medical Research Center, School of Basic Medical Science, Fudan University lhqu20@fudan.edu.cn Xiaoyuan Luo Digital Medical Research Center, School of Basic Medical Science, Fudan University 19111010030@fudan.edu.cn Manning Wang Digital Medical Research Center, School of Basic Medical Science, Fudan University mnwang@fudan.edu.cn Zhijian Song Digital Medical Research Center, School of Basic Medical Science, Fudan University zjsong@fudan.edu.cn
Pseudocode No The paper describes its methods using prose and mathematical equations but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/miccaiif/WENO.
Open Datasets Yes We used five datasets to comprehensively evaluate the performance of WENO, including two synthetic datasets and three real-world datasets. To explore the performance of WENO under different positive instance ratios, we used the 10-class natural image dataset CIFAR 10 [15] and the 9-class pathological image dataset CRC [14] to construct synthetic WSI datasets with different positive instance ratios... Furthermore, we used real-world pathology datasets from three different medical centers to evaluate the performance of WENO, including a breast cancer lymph node metastasis public dataset, the Camelyon16 dataset [1], a lung cancer diagnosis public dataset, the TCGA Lung Cancer dataset, and an in-house cervical cancer lymph node metastasis dataset, the Clinical Cervical dataset.
Dataset Splits No For the hard positive instance mining strategy, we drop the instances with positive probability higher than a threshold in positive bags. The hyperparameter thresholds vary for each dataset, and we used grid search on the validation set to determine the optimal values. While a 'validation set' is mentioned, the paper does not specify the exact split percentages or absolute sample counts for the training, validation, and test sets to reproduce the data partitioning.
Hardware Specification Yes All experiments were performed using 4 Nvidia 3090 GPUs.
Software Dependencies No The paper does not specify software dependencies with version numbers (e.g., 'Python 3.x', 'PyTorch 1.x'). It only mentions general tools like 'SGD optimizer'.
Experiment Setup Yes For the CIFAR-10-MIL dataset, the encoder in Figure 2 is implemented using the Alex Net [16]. For the other datasets, the encoder is implemented using the Res Net18 [10]. Both the prediction heads and the attention module are implemented using fully connected layers. No pre-training of the network parameters and no image augmentation are performed. The SGD optimizer is used to optimize the network parameters with a fixed learning rate of 0.001. For the hard positive instance mining strategy, we drop the instances with positive probability higher than a threshold in positive bags. The hyperparameter thresholds vary for each dataset, and we used grid search on the validation set to determine the optimal values.