Few Shot Learning Geometric Ensemble for Multi-label Chest X-Rays

Authors: Dana Moukheiber*, Saurabh Mahindre*, Lama Moukheiber, Mira Moukheiber, Song Wang, Chunwei Ma, George Shih, Yifan Peng, and Mingchen Gao

(*) Equal contribution

Paper: Link

Dataset

The dataset labels are available in the labels folder.

MIMIC CXR dataset and corresponding images are available at: physionet.org/content/mimic-cxr/2.0.0/ . You need to sign-up as a user on physionet and sign the data use agreement.

Model Weights

The weights for pretrained base feature extractor (ResNet) model, NCA model and feature statistics used in Distribution Calibration are available here: Google Drive Link

Notebooks

• Deepvoro.ipynb: Jupyter notebook with code for:

  • Loading model weights
  • DC Voronoi LR
  • NCA Loss finetuning
  • BCE Loss finetuning
  • Episodic evaluation
  • Finetuning
  • Deepvoro ensemble
  • Few-shot evaluation

• few_shot_cxray: Module with code for:

  • Residual Network backbone
  • Multilabel NCA Loss definition
  • Dataset loaders
  • Utilities

Citation

Please cite our work if you find it useful!

@InProceedings{10.1007/978-3-031-17027-0_12,
author="Moukheiber, Dana
and Mahindre, Saurabh
and Moukheiber, Lama
and Moukheiber, Mira
and Wang, Song
and Ma, Chunwei
and Shih, George
and Peng, Yifan
and Gao, Mingchen",
editor="Nguyen, Hien V.
and Huang, Sharon X.
and Xue, Yuan",
title="Few-Shot Learning Geometric Ensemble for Multi-label Classification of Chest X-Rays",
booktitle="Data Augmentation, Labelling, and Imperfections",
year="2022",
publisher="Springer Nature Switzerland",
address="Cham",
pages="112--122",
abstract="This paper aims to identify uncommon cardiothoracic diseases and patterns on chest X-ray images. Training a machine learning model to classify rare diseases with multi-label indications is challenging without sufficient labeled training samples. Our model leverages the information from common diseases and adapts to perform on less common mentions. We propose to use multi-label few-shot learning (FSL) schemes including neighborhood component analysis loss, generating additional samples using distribution calibration and fine-tuning based on multi-label classification loss. We utilize the fact that the widely adopted nearest neighbor-based FSL schemes like ProtoNet are Voronoi diagrams in feature space. In our method, the Voronoi diagrams in the features space generated from multi-label schemes are combined into our geometric DeepVoro Multi-label ensemble. The improved performance in multi-label few-shot classification using the multi-label ensemble is demonstrated in our experiments (The code is publicly available at https://github.com/Saurabh7/Few-shot-learning-multilabel-cxray).",
isbn="978-3-031-17027-0"
}