# Review of CNN-Based Approaches for Preprocessing, Segmentation and Classification of Knee Osteoarthritis

## TL;DR

This paper reviews deep learning methods for diagnosing knee osteoarthritis, comparing their accuracy and limitations.

## Contribution

The paper systematically reviews and compares CNN-based approaches for knee osteoarthritis classification, highlighting performance and limitations.

## Key findings

- CNN-based methods for KOA classification achieve accuracies ranging from 61% to 98%.
- X-ray and MRI datasets are commonly used, with performance varying by imaging modality.
- The paper identifies methodological limitations and suggests future research directions for more robust systems.

## Abstract

Osteoarthritis (OA) is a prevalent joint disorder characterized by symptoms such as pain and stiffness, often leading to loss of function and disability. Knee osteoarthritis (KOA) represents the most prevalent type of osteoarthritis. KOA is usually detected using X-ray radiographs of the knee; however, the classification of disease severity remains subjective and varies among clinicians, motivating the need for automated assessment methods. In recent years, deep learning–based approaches have shown promising performance for KOA classification tasks, particularly when applied to structured imaging datasets. This review analyzes convolution neural network (CNN)-based approaches reported in the literature and compares their performance across multiple criteria. Studies were identified through systematic searches of IEEE Xplore, SpringerLink, Elsevier (ScienceDirect), Wiley Online Library, ACM Digital Library, and other sources such as PubMed and arXiv, with the last search conducted in March 2025. The review examines datasets used (primarily X-ray and MRI), preprocessing strategies, segmentation techniques, and deep learning architectures. Reported classification accuracies range from 61% to 98%, depending on the dataset, imaging modality, and task formulation. Finally, this paper highlights key methodological limitations in existing studies and outlines future research directions to improve the robustness and clinical applicability of deep learning–based KOA classification systems.