Hi-DER

This is the official PyTorch implementation of our La Radiologia Medica publication:
“Classification models for arthropathy grades of multiple joints based on hierarchical continual learning.” (2025)

We propose Hierarchical Dynamically Expandable Representation (Hi-DER), a hierarchical continual arthropathy classification model that can be continuously updated to arthropathy classification of multiple joints, by using a continual learning strategy. Please refer to our paper for more details.

Table of Contents

• Introduction
• Proposed Architecture
• DeLong's Test
• Model Robustness: Addressing Underrepresented Features of Osteoarthritis
• Pre-requisites
• Hyperparameters
• Run Experiment
• Acknowledgement
• Citation

Introduction

Although arthropathy can affect diverse joints with different morphologies and pathological changes, most of the previous studies are designed for single joint classification with limited generalizability. To date, there has been no study that developed a multiple joint arthropathy classification model with continuously expandable classification capabilities. The proposed hierarchical continual classification model can be continuously updated for large-scale studies of multiple joints with various anatomical structures.

Proposed Architecture

DeLong's Test

The Hi-DER was developed using knee, elbow, ankle, and shoulder radiographs from Sinchon Severance hospital. To evaluate whether utilizing large public datasets affects the predictive performance of Hi-DER, we conducted DeLong's test comparing the AUCs of the model that employed the pre-trained feature extractor and the original model which was trained solely on internal datasets. The pre-training was done using the Osteoarthritis Initiative dataset. Below, we provide the p-values from DeLong’s test at a significance level of 0.05, showing no significant differences in most cases.

Model Robustness: Addressing Underrepresented Features of Osteoarthritis

The model's performance can be influenced by the underrepresentation of rare and mild cases. For example, ankle valgus osteoarthritis is significantly less common than varus osteoarthritis, leading to reduced diagnostic accuracy for such atypical cases. Additionally, early osteoarthritis with subtle minor osteophytes often poses challenges for osteoarthritis detection algorithms. These limitations highlight the need for more diverse and balanced datasets to enhance the model performance and generalizability. Our approach accounts for these challenges by focusing on generalized pathological findings across joint types such as joint space narrowing and osteophyte formation.

Pre-requisites

Run the following commands to clone this repository and install the required packages.

git clone https://github.com/DigitalHealthcareLab/24HiDER.git
pip install -r requirements.txt

Hyperparameters

• memory_size: The total number of preserved exemplar in the incremental learning process.
• memory_per_class: The number of preserved exemplar per class ($\frac{memory size}{K}$).
• shuffle: Whether to shuffle the class order or not.
• init_cls: The number of classes in the initial incremental step.
• increment: The number of classes in each incremental step $t$ ($t$ > 1).

Run Experiment

• Edit the hider.json file for global settings.
• Set hyperparameters for training in the hider.py file.
• Set data path in the data.py file.
• Edit the structure of hierarchical labels in the base.py, hider.py, and hierarchical_loss.py file.
• Run the following command to run the experiment.

python main.py --config=./exps/hider.json

Acknowledgement

Our code is based on PyCIL. We thank the authors for providing the great base code.

Citation

If you find this code useful, please consider citing our paper:

@article{jang2025classification,
  title={Classification models for arthropathy grades of multiple joints based on hierarchical continual learning},
  author={Jang, Bong Kyung and Kim, Shiwon and Yu, Jae Yong and Hong, JaeSeong and Cho, Hee Woo and Lee, Hong Seon and Park, Jiwoo and Woo, Jeesoo and Lee, Young Han and Park, Yu Rang},
  journal={La radiologia medica},
  pages={1--13},
  year={2025},
  publisher={Springer}
}