LEE- Bilgisayar Mühendisliği-Yüksek Lisans

Bu koleksiyon için kalıcı URI

http://hdl.handle.net/11527/19207

Gözat

Yazar "Aktemur, İlknur" ile LEE- Bilgisayar Mühendisliği-Yüksek Lisans'a göz atma
  • Öge
    Classification of knee osteoarthritis severity using deep learning with fully supervised and semi-supervised-based approaches
    (Graduate School, 2024-07-02) Aktemur, İlknur ; Öksüz, İlkay ; 504201575 ; Computer Engineering
    Osteoarthritis (OA) occurs when the cartilage tissue, a smooth substance found at the ends of bones, thins over time and even disappears. Symptoms such as swelling, pain, and restricted movement in the affected knees significantly diminish the quality of life for individuals and pose a significant health problem that limits their daily activities. To alleviate the damages caused by osteoarthritis and slow down the progression of the disease, accurate determination of the disease's stages is essential. Cartilage degeneration, wear, reduced joint space, and bone changes seen on X-rays offer indirect insights into knee OA stages. This situation causes disagreement among radiologists and hinders early detection. Recently, there has been an increase in studies focusing on accelerating early diagnosis and supporting radiologists' decisions by automatically determining disease stages with deep learning-based methods. However, high-performance models are still needed. The success of deep learning often depends on large-scale labeled data. However, the data labeling phase of medical images is costly and laborious. Within the scope of this thesis, we recognized this challenge, we proposed fully supervised and semi-supervised methods suited for scenarios with abundant or limited labeled data. Firstly, we fine-tuned models like ResNet, DenseNet, EfficientNet,ResNeXt, ConvNeXt trained on the ImageNet using the publicly available OAI dataset in fully supervised learning part. Also during fine-tuning, we used the categorical cross-entropy loss function and the ordinal loss function. We called this process one-stage fine-tuning. Fine-tuning with the ordinal loss function yielded the best results on the OAI test data, with a 73.91% accuracy score and a 73.77% F1-score achieved by the ConvNeXt V2-Tiny model. The ConvNeXt V1-Base model provided the best result in Mean Absolute Error (MAE) metric with a score of 0.28. For scenario with a limited amount of labeled data gathered from a single center, we employ a two-stage fine-tuning approach within the realm of fully supervised methods. We re-fine-tune models that obtained from first stage-fine tuning using smaller-scale hospital data this time. The two-stage fine-tuning method has yielded higher performance scores across all models. The best model performance belongs to the ConvNeXt V1-Base model again, with a 70.86% accuracy score, 71.35% F1-score, and a 0.30 MAE score. For scenarios where unlabeled data more than labeled data, we propose a semi-supervised approach utilizing both types of data. Here, we integrate representations learned through a pre-training process based on Contrastive Learning(CL) using unlabeled X-ray data into a supervised fine-tuning process conducted with a small amount of labeled data. We also examine model robustness by employing semi-supervised learning, gradually reducing labeled data by 25% in each experiment. Our model trained with the semi-supervised learning approach outperforms the baseline model with a 64.16% accuracy score, 64.72% F1 score, and 0.42 MAE score. Furthermore, as the labeled data decreases, the gap between them widens for all performance metrics.