IMAML-IDCG: Optimization-Based Meta-Learning with ImageNet Feature Reusing for Few-Shot Invasive Ductal Carcinoma Grading

This repository contains the reference source code for the paper IMAML-IDCG: Optimization-Based Meta-Learning with ImageNet Feature Reusing for Few-Shot Invasive Ductal Carcinoma Grading.

Abstract

Model-Agnostic Meta-Learning (MAML) is a widely used few-shot learning (FSL) technique that reduces reliance on large, labeled datasets in deep learning for medical imaging analysis. However, MAML requires backpropagating through all feature layers for task adaptation, leading to suboptimal computational efficiency. We propose IMAML-IDCG (ImageNet Model-Agnostic Meta-Learning in Invasive Ductal Carcinoma Grading), which enhances computational efficiency for few-shot grading of Invasive Ductal Carcinoma (IDC) through three key techniques:

1. ImageNet Feature Reusing: IMAML-IDCG is initialized with ImageNet pre-trained weights, allowing the reuse of rich feature representations learned from the large-scale ImageNet dataset.
2. ImageNet Partial Freezing Strategy: During the inner optimization loop, only the model’s classifier head layer is optimized, while the last few layers of the model are selectively fine-tuned in the outer loop to enhance efficiency and reduce overfitting.
3. Adaptive Inner Learning Rate: An adaptive learning rate is employed during task adaptation to improve convergence and performance.

We evaluated IMAML-IDCG using the BreaKHis dataset (7,909 images) as the base dataset, and the BCHI (282 images) and PathoIDCG (3,744 images) datasets as the novel datasets. Our empirical results demonstrate that IMAML-IDCG outperforms MAML and other FSL methods in few-shot IDC grading tasks across various cross-magnification domain settings. Notably, IMAML-IDCG achieves a 14.64% improvement over MAML on the BCHI dataset and a 6.04% improvement on the PathoIDCG 40X dataset when meta-trained with the BreaKHis 40X dataset in the 3-way 5-shot scenario.

Citation

If you find our code useful, please consider citing our work using the following BibTeX entry:

@article{voon2024imamlidcg,
  title={IMAML-IDCG: Optimization-based meta-learning with ImageNet feature reusing for few-shot invasive ductal carcinoma grading},
  author={Wingates Voon, Yan Chai Hum, Yee Kai Tee, Wun-She Yap, Khin Wee Lai, Humaira Nisar, Hamam Mokayed},
  journal={Expert Systems with Applications},
  volume={257},
  pages={124969},
  year={2024},
  publisher={Elsevier},
  issn={0957-4174},
  doi={https://doi.org/10.1016/j.eswa.2024.124969},
  url={https://www.sciencedirect.com/science/article/pii/S0957417424018360}
}

Enviroment

• Google Colab
• Google Drive
• Python3
• Pytorch
• json

Getting started

Clone the Repo

• Clone the repo into your Google Colab working directory

!git clone https://github.com/wingatesv/IMAML-IDCG.git

Datasets Download

• Please contact the author for more information: [email protected]

BreaKHis Class	4X	10X	20X	40X
Adenosis	114	113	111	106
Phyllodes Tumor	149	150	140	130
Tubular Adenoma	109	121	108	115
Lobular Carcinoma	156	170	163	137
Papillary Carcinoma	145	142	135	138
Mucinous Carcinoma	205	222	196	169
Fibroadenoma	253	260	264	237
Ductal Carcinoma	864	903	896	788
Total	1995	2081	2013	1820

BCHI Class	40X
Grade 1	98
Grade 2	102
Grade 3	82

PathoIDC Class	20X	40X
Grade 1	600	317
Grade 2	641	480
Grade 3	1245	361

Split	Dataset	link
Base (BC Classification)	BreaKHis_4x	https://drive.google.com/file/d/1CWRzSFRV1WsRyA_g8yYQvIQGs_QnUyYR/view?usp=sharing
	BreaKHis_10x	https://drive.google.com/file/d/1-5PuA8i4vmUN1fHAH8pveEjhLn5S5ACU/view?usp=sharing
	BreaKHis_20x	https://drive.google.com/file/d/1-JcLWy-cFmd7LVdcKaXPwJK-ox3_r5m2/view?usp=share_link
	BreaKHis_40x	https://drive.google.com/file/d/1yDbX4Nm39zOag4jqLeoxeWSQ3SeAZ0Aa/view?usp=share_link
Novel (IDC Grading)	BCHI	https://drive.google.com/file/d/1ssUKwT2jByxMPQahS58TFH_nga2yQ9D1/view?usp=share_link
	PathoIDC_20x	https://drive.google.com/file/d/1dXuY1TePiwcrsrvFW-I3-8CidaVXmgMx/view?usp=share_link
	PathoIDC_40x	https://drive.google.com/file/d/1-J6Hw2MZP5Grnqfgpl1-ZNk_pCGJWtyI/view?usp=share_link

BreaKHis_4x

• Change directory to ./filelists/BreaKHis_4x
• run source ./get_BreaKHis_4x.sh

BreaKHis_10x

• Change directory to ./filelists/BreaKHis_10x
• run source ./get_BreaKHis_10x.sh

BreaKHis_20x

• Change directory to ./filelists/BreaKHis_20x
• run source ./get_BreaKHis_20x.sh

BreaKHis_40x

• Change directory to ./filelists/BreaKHis_40x
• run source ./get_BreaKHis_40x.sh

BCHI

• Change directory to ./filelists/BCHI
• run source ./get_BCHI.sh

PathoIDC_20x

• Change directory to ./filelists/PathoIDC_20x
• run source ./get_PathoIDC_20x.sh

PathoIDC_40x

• Change directory to ./filelists/PathoIDC_40x
• run source ./get_PathoIDC_40x.sh

BreaKHis_{}X->BCHI

• Finish preparation for BreaKHis and BCHI and you are done!

BreaKHis_{}x->PathoIDC{}x

• Finish preparation for BreaKHis and PathoIDC and you are done!

Self-defined setting

• • Require one data split json file: 'base.json' for each BreaKHis dataset
• • Require two data split json file: 'val.json', 'novel.json' for BCHI and PathoIDC datasets
• The format should follow
  {"label_names": ["class0","class1",...], "image_names": ["filepath1","filepath2",...],"image_labels":[l1,l2,l3,...]}
• For example, run source ./get_BCHI_filelist.py to generate json files for the BCHI dataset
• Put these file in the same folder and change data_dir['DATASETNAME'] in configs.py to the folder path

Cross-Domain Configurations

Name	Description
cross_IDC_4x	BreaKHis_4x to BCHI
cross_IDC_10x	BreaKHis_10x to BCHI
cross_IDC_20x	BreaKHis_20x to BCHI
cross_IDC_40x	BreaKHis_40x to BCHI
cross_IDC_4x_2	BreaKHis_4x to PathoIDC_40x
cross_IDC_10x_2	BreaKHis_10x to PathoIDC_40x
cross_IDC_20x_2	BreaKHis_20x to PathoIDC_40x
cross_IDC_40x_2	BreaKHis_40x to PathoIDC_40x
cross_IDC_4x_3	BreaKHis_4x to PathoIDC_20x
cross_IDC_10x_3	BreaKHis_10x to PathoIDC_20x
cross_IDC_20x_3	BreaKHis_20x to PathoIDC_20x
cross_IDC_40x_3	BreaKHis_40x to PathoIDC_20x

Train

Run python ./train.py --dataset [DATASETNAME] --model [BACKBONENAME] --method [METHODNAME] [--OPTIONARG]

For example, run python ./train.py --dataset cross_IDC_40x --model ResNet34 --method imaml_idcg --train_n_way 3 --test_n_way 3 --n_shot 1 --stop_epoch 100 --train_aug --sn stainnet
Commands below follow this example, and please refer to io_utils.py for additional options.

Save features

Save the extracted feature before the classifaction layer to increase test speed. This is not applicable to MAML-based methods, but are required for other methods. Run python ./save_features.py --dataset cross_IDC_40x --model ResNet34 --method relationnet --train_n_way 3 --n_shot 5 --test_n_way 3 --train_aug --sn stainnet

Test

Run python ./test.py --dataset cross_IDC_40x --model ResNet34 --method imaml_idcg --train_n_way 3 --test_n_way 3 --n_shot 1 --train_aug --sn stainnet

Results

• The test results will be recorded in ./record/results.txt

References

• Main Framework https://github.com/wyharveychen/CloserLookFewShot
• Framework, Backbone, Method: Matching Network https://github.com/facebookresearch/low-shot-shrink-hallucinate
• Omniglot dataset, Method: Prototypical Network https://github.com/jakesnell/prototypical-networks
• Method: Relational Network https://github.com/floodsung/LearningToCompare_FSL
• Method: MAML https://github.com/cbfinn/maml
  https://github.com/dragen1860/MAML-Pytorch
  https://github.com/katerakelly/pytorch-maml