CXRLT24 Multiview PP

This repository contains our solution for the MICCAI 2024 CXR-LT (Chest X-Ray Long-Tailed) challenge, achieving 4th place in Subtask 2 and 5th in Subtask 1.

Project Overview

We present an ensemble method for long-tailed chest X-ray (CXR) classification using ConvNeXt V2 and MaxViT models. Our approach combines state-of-the-art image classification techniques with asymmetric loss for handling class imbalance and view-based prediction aggregation to enhance overall performance.

Repository Structure

The code directory contains the following Python scripts:

• config.py: Configuration settings for the project
• dataset.py: Dataset handling and preprocessing
• inference.py: Model inference logic
• model.py: Model architecture definitions
• postprocess.py: Post-processing techniques including view-based aggregation
• run_inference.py: Script to run inference on test data
• run_training.py: Script to initiate the training process
• train.py: Training loop and logic
• utils.py: Utility functions used across the project

Citation

If you find this work useful for your research, please consider citing our paper:

@misc{yamagishi2024ensembleconvnextv2maxvit,
      title={Ensemble of ConvNeXt V2 and MaxViT for Long-Tailed CXR Classification with View-Based Aggregation}, 
      author={Yosuke Yamagishi and Shouhei Hanaoka},
      year={2024},
      eprint={2410.10710},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2410.10710}, 
}

A summary paper for the MICCAI 2024 CXR-LT Challenge will be published in the future. Once available, we will update this section with the relevant citation information.