Date of Award

2025

Document Type

Thesis

Degree Name

Master of Science in Artificial Intelligence

Department

Computer Science

Committee Chair and Members

Reda Nacif Elalaoui, Chair

Abla Bedoui

Debarshi Ghosh

Keywords

Edge detection, Image segmentation, KL grading, Knee osteoarthritis (KOA) classification, Medical imaging, YOLOv8

Abstract

Knee Osteoarthritis (KOA) is a degenerative joint condition characterized by the progressive narrowing of joint space and structural deterioration. The structural degradation of the joint space is evaluated using the Kellgren–Lawrence (KL) grading system, and accurate classification across all grades, specifically in the early stages, remains a challenge owing to subtle radiographic differences. This study presents an automated KL-grade classification framework that integrates joint edge enhancement with deep learning to improve KOA grading using radiographic images.

Edge detection filters, namely Sobel, Scharr, and Canny, were applied to X-ray images to enhance the joint space boundaries and osteoarthritic features. These preprocessed images were used to train a YOLOv8 model capable of simultaneously predicting KL grades and segmenting the joint regions. The combined output of the bounding boxes and segmentation masks enabled the model to link anatomical features with disease severity in a single architecture.

The model performance was evaluated using precision, recall, mean Average Precision (mAP), and F1-score across all KL grades. The Scharr-enhanced model achieved the best overall detection results (box mAP@0.5 = 0.679; mask mAP@0.5:0.95 = 0.300), particularly for the KL-2 and KL-3 grades. The Sobel-enhanced model attained the highest recall for KL-1 (0.692) and superior segmentation for KL-4 (mask mAP@0.5 = 0.709), showing strong performance in the early and severe stages. In contrast, the Canny-based model exhibited reduced segmentation accuracy due to sparse and discontinuous edge patterns. The findings highlight that the edge-enhanced YOLOv8 framework improves the classification and segmentation of mid-to severe KOA stages. The proposed approach demonstrates that integrating edge filtering with object detection contributes to a more anatomically grounded KL-grade classification while emphasizing the need for further optimization to improve sensitivity in early-stage KOA.

Keywords: Knee Osteoarthritis (KOA), KL grade, YOLO V8, Sobel filter, Canny Edge Detector, Mean Average Precision (mAP)

Download

Share

COinS