[Paper Review] An Application of Backpropagation Artificial Neural Network Method for Measuring The Severity of Osteoarthritis

Dian Pratiwi, Diaz D. Santika+1 more|arXiv (Cornell University)|Sep 29, 2013

Medical Imaging and AnalysisEngineering3 references16 citations

TL;DR

This study proposes a backpropagation artificial neural network (BP-ANN) system to classify osteoarthritis severity from hand and wrist X-ray images. The method uses image preprocessing, statistical texture features, and 132 normalized input vectors to achieve 100% accuracy on training data and 80% on combined learning/non-learning data, demonstrating potential for automated osteoarthritis severity assessment.

ABSTRACT

The examination of Osteoarthritis disease through X-ray by rheumatology can be classified into four grade of severity. This paper discusses about the application of artificial neural network backpropagation method for measuring the severity of the disease, where the observed X-ray range from wrist to fingers. The main procedures of system in this paper is divided into three, which are image processing, feature extraction, and artificial neural network process. First, an X-ray image digital (200x150 pixels and greyscale) will be thresholded, then extracted features based on probabilistic values of the color intensity of seven bit quantization result, and statistical textures. That feature values then will be normalizing to interval [0.1, 0.9], and then the result would be processing on backpropagation artificial neural network system as input to determine the severity of disease from an X-ray had input before it. From testing with learning rate 0.3, momentum 0.4, hidden units five pieces and about 132 feature vectors, this system had had a level of accuracy of 100% for learning data, 80% for learning and non-learning data, and 66.6% for non-learning data

Motivation & Objective

• To develop an automated system for assessing osteoarthritis severity using medical imaging and machine learning.
• To address the subjectivity and variability in radiological grading of osteoarthritis by rheumatologists.
• To apply artificial neural networks to improve consistency and objectivity in disease severity classification.
• To evaluate the performance of a backpropagation neural network on X-ray images of the hand and wrist.
• To explore the feasibility of using texture and intensity-based features for osteoarthritis severity prediction.

Proposed method

• X-ray images (200×150 pixels, grayscale) were preprocessed through thresholding to enhance contrast.
• Feature extraction involved computing probabilistic intensity values from seven-bit quantized pixel intensities and statistical texture features.
• Extracted features were normalized to the range [0.1, 0.9] to improve neural network convergence.
• A backpropagation neural network with five hidden units was trained using a learning rate of 0.3 and momentum of 0.4.
• The system used 132 feature vectors as input to predict osteoarthritis severity across four grades.
• The network was trained and tested on both learning and non-learning data subsets to evaluate generalization.

Experimental results

Research questions

• RQ1 Can a backpropagation neural network accurately classify osteoarthritis severity from hand and wrist X-ray images?
• RQ2 How does the system perform on training data versus unseen (non-learning) test data?
• RQ3 What is the impact of feature normalization and texture-based input on classification accuracy?
• RQ4 To what extent can an ANN reduce subjectivity in radiological assessment of osteoarthritis?
• RQ5 How do learning rate and momentum settings affect the model's convergence and performance?

Key findings

• The system achieved 100% accuracy on the training dataset, indicating strong learning capability.
• The model reached 80% accuracy when tested on both learning and non-learning data, suggesting moderate generalization.
• On non-learning data alone, the system achieved 66.6% accuracy, indicating some overfitting or limited generalization to new cases.
• Feature normalization to [0.1, 0.9] improved training stability and convergence of the backpropagation algorithm.
• The use of seven-bit intensity quantization and statistical texture features provided sufficient discriminative input for the ANN.
• The model with five hidden units and optimized learning parameters demonstrated feasibility for automated osteoarthritis severity classification.