A Multiscale Frequency Domain Causal Framework for Enhanced Pathological Analysis

Authors: Xiaoyu Cui, Weixing Chen, Jiandong Su

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on Camelyon16 and TCGANSCLC dataset show that, compared to previous work, our method has significantly improved accuracy and generalization ability, providing a new theoretical perspective for medical image analysis and potentially advancing the field further. 4 EXPERIMENT
Researcher Affiliation Academia 1Northeastern University 2 Sun Yat-sen University 3Shenzhen Institute of Advanced Technology EMAIL, EMAIL, EMAIL
Pseudocode No The paper describes methods and modules (CMIM, MSRM, FSRM) but does not present them in a structured pseudocode or algorithm block format.
Open Source Code Yes The code will be released at https://github.com/Wissing Chen/ MFC-MIL.
Open Datasets Yes The Camelyon16 dataset Bejnordi et al. (2017) is widely used for detecting breast cancer metastases. ... Meanwhile, the TCGA-NSCLC dataset focuses on two lung cancer subtypes, LUSC and LUAD, with 1,054 whole slide images.
Dataset Splits Yes The Camelyon16 dataset... 270 training and 129 testing images... Meanwhile, the TCGA-NSCLC dataset... It is divided into training, validation, and test sets in a 7:1:2 ratio... To evaluate the effectiveness of our approach, we apply four key metrics for classification performance: accuracy, F1 score, specificity, and the area under the receiver operating characteristic curve (AUC). These metrics provide a comprehensive assessment of the method s overall performance... using 5-fold cross-validation.
Hardware Specification Yes All experiments were conducted on an NVIDIA Ge Force RTX 2080Ti.
Software Dependencies No In the feature extraction process, we employed a CNN-based Res Net18, with parameters pre-trained using Sim CLR as part of the DSMIL framework. ... we used the Adam optimizer with an initial learning rate of 2e-4 and a weight decay of 5e-4.
Experiment Setup Yes The model operates with a dimension of 512, while the value of k in CMIM is set to 16 for high-resolution features and 32 for low-resolution features. For most experiments, we used the Adam optimizer with an initial learning rate of 2e-4 and a weight decay of 5e-4. Additionally, our MFC estimates the mediator using patch-level features and applies it to intervene in the aggregated bag-level prediction vector. The mini-batch size used for training is 1, and the model is trained for 100 epochs.