Deep Convolutional Neural Network-Based Automatic Classification of Ovarian Sections With Different Histological Staining Techniques
16th INTERNATIONAL ZEUGMA CONFERENCE ON SCIENTIFIC RESEARCHES, Gaziantep, Türkiye, 26 - 28 Haziran 2026, ss.138-139, (Özet Bildiri)
- Yayın Türü: Bildiri / Özet Bildiri
- Basıldığı Şehir: Gaziantep
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.138-139
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Kayseri Üniversitesi Adresli: Evet
Özet
Accurate and reliable classification of ovarian images is considered a fundamental requirement for the development of clinical decision support systems and the rigorous monitoring of treatment processes. Identifying structural changes in ovarian tissue, particularly those caused by drug use, is critical for evaluating treatment efficacy and creating patient-specific treatment plans. This study presents a deep learning-based classification approach for analyzing ovarian follicle images from female rats. The histological sections used in this study were prepared using Hematoxylin-Eosin (H&E), Anti-Müllerian Hormone (AMH) immunohistochemical staining, and Masson trichrome techniques to highlight different morphological features of tissue and cellular structures. This comprehensive dataset, generated from experimental processes, underwent a series of preprocessing steps prior to the classification stage to improve image quality and optimize model performance. Accordingly, image cropping was applied, and data-balancing methods were used to address class imbalance. These preparation processes resulted in a dataset that is suitable and standardized for deep learning analyses. In the classification phase, the transfer learning strategy was adopted, and the ResNet50, EfficientNetB0, and InceptionResNetV2 architectures were trained. The performance of these models was then analyzed comparatively. Academically accepted metrics, including accuracy, precision, sensitivity, F1 score, macro-average F1 score, and ROC-AUC, were used to evaluate the models' performance. Examining the experimental results, it was determined that the ResNet50 model achieved the highest overall classification performance, with 66.94% accuracy and a macro F1 score of 65.91%. The EfficientNetB0 model achieved superior inter-class discrimination, with an ROC-AUC of 0.9115. In contrast, the InceptionResNetV2 architecture performed worse than the other two models. These findings confirm that structural similarities between classes make the classification of over images a complex problem. However, deep learning models have been found to be competent in learning distinctive features despite this complexity. This study highlights the potential of deep learning approaches for automated analysis of drug-induced ovarian changes and provides a scientific basis for future research in digital pathology. These data, which reveal the effects of different staining techniques on model performance, serve as an important reference for the development of clinical decision support mechanisms.