Deep Learning-Based Classifıcation of Ovarian Images For Cancer Detection
Congress on Innovative Approaches in Natural and Applied Sciences, Erzurum, Türkiye, 20 - 22 Nisan 2026, ss.439-449, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Basıldığı Şehir: Erzurum
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.439-449
- Kayseri Üniversitesi Adresli: Evet
Özet
Early and highly accurate classification of ovarian images is a vital research area. Developing computer-aided systems can directly contribute to ovarian cancer diagnosis processes. This study addresses a comprehensive deep learning-based classification methodology. It automatically classifies complex ovarian images into two categories: "cancerous" and "healthy." For the experimental basis, a comprehensive ovarian image dataset from Kaggle, accessible to researchers, was chosen. The obtained data was meticulously split into three subsets—training, validation, and testing. This approach objectively assesses the models' learning capacity and generalization. The classification architecture uses transfer learning, a strategy that has proven successful in the current literature. Accordingly, three different convolutional neural network (CNN) architectures—ResNet50, InceptionResNetV2, and EfficientNetB0—which exhibit high performance in object recognition and medical image analysis, were included in the process. The models' classification capabilities were subjected to a detailed comparative analysis using multidimensional performance metrics, including accuracy, precision, recall, F1-score, and ROC-AUC. The data obtained from the experimental studies show that all three architectures exhibit high success rates in distinguishing medical images. However, the evaluations showed that the ResNet50 model achieved the strongest and most consistent performance. The ResNet50 architecture outperformed the other models, achieving accuracy of 0.9758, precision of 0.9788, recall of 0.9808, and F1-score of 0.9798. Furthermore, the ROC-AUC value, which reflects the model's discriminatory power, was very high at 0.9947. Comparatively, the InceptionResNetV2 model achieved an accuracy of 0.9660 and an F1 score of 0.9716, while the EfficientNetB0 model remained at an accuracy of 0.9620 and an F1 score of 0.9684. The findings clearly demonstrate that deep learning models trained with a transfer learning approach, particularly the ResNet50 architecture, have the potential to guide clinicians in ovarian cancer diagnosis. The deep feature representation capabilities offered by pre-trained networks have proven to be a fundamental component of clinical decision support systems, delivering high generalization performance even on limited datasets. This study supports that provides important data supporting the reliability and applicability of AIbased diagnostic tools in healthcare.