An Explainable Patch-Based Deep Learning Framework for Marker-Specific Computational Assessment of Ovarian Autophagy in Digital Histopathology
Bioengineering, cilt.13, sa.1103, ss.1-20, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 13 Sayı: 1103
- Basım Tarihi: 2026
- Doi Numarası: 10.3390/bioengineering13101103
- Dergi Adı: Bioengineering
- Derginin Tarandığı İndeksler: Academic Search Ultimate (EBSCO), Natural Science Collection (ProQuest), Biological Science Database (ProQuest), Scopus, Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest), Science Citation Index Expanded (SCI-EXPANDED), BIOSIS, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.1-20
- Kayseri Üniversitesi Adresli: Evet
Özet
Autophagy is essential for ovarian physiology and is implicated in pathological conditions such as Polyendocrine Metabolic Ovarian Syndrome (PMOS). The evaluation of autophagy-related markers in histopathological images is often subjective and observer-dependent. This study presents an explainable computational pathology framework that integrates patch-based deep learning with Grad-CAM-based visual interpretation to enable objective and reproducible analysis of autophagy-related immunohistochemical (IHC) staining in rat ovarian tissue. Beclin-1, LC3, and p62-stained rat ovarian tissue images from control, PMOS, bee bread (PERGA), and PERGA + PMOS groups were analyzed using a patch-based deep learning approach. Multiple convolutional neural network architectures, including DenseNet121, EfficientNetV2B0, and ConvNeXtTiny, were trained and evaluated, and their outputs were combined through an ensemble strategy. Image-level predictions were generated by averaging patch-level probabilities, and Grad-CAM was used to visualize image regions influencing model decisions. The LC3-based ensemble model achieved the highest overall image-level classification performance, with an F1-score of 78.28% and an AUC of 88.15%. In contrast, the Beclin-1 DenseNet121 model demonstrated the highest discrimination performance based on AUC, reaching 93.63%. Models trained on p62 images showed lower classification performance, likely due to the heterogeneous staining characteristics of p62 expression. Grad-CAM visualizations revealed that model activation was predominantly localized to DAB-positive and tissue-informative regions, supporting the interpretability of the proposed framework. In summary, this approach enables objective, reproducible, and interpretable analysis of autophagy-related histopathological images. The findings suggest that patch-based deep learning combined with explainable artificial intelligence offers a promising computational tool for assessing autophagy-related alterations in rat ovarian tissue.