A Pretrained CNN–Transformer Gated Fusion Architecture for Ovarian Cancer Histological Subtype
Document Type
Conference Proceeding
Publication Date
2026
Abstract
Ovarian cancer is considered to be among the most lethal gynecologic malignancies, as it has a predominantly asymptomatic manifestation causing late diagnoses. The histological subtype identification based on the Hematoxylin and Eosin (H&E) stained whole slide images is essential to personalized treatment but is a difficult task due to inter-observer variation. In this paper, OvarianHybrid is introduced: a gated Convolutional Neural Network (CNN)-Transformer architecture in which input patches first pass through a controlled preprocessing pipeline, followed by a pretrained DenseNet121 backbone whose spatial feature tokens are fed into a randomly initialized four-block Transformer, and the two streams are fused via a learnable sigmoid gate for final classification into five histological subtypes. A structured three-phase ablation is conducted on a public Kaggle dataset of 34,266 H&E patches to independently analyze the contributions of preprocessing, class imbalance handling, and data augmentation. Grad-CAM visual explanations are provided to support clinical interpretability. OvarianHybrid achieves 9 5. 5 5 % accuracy and 9 5. 5 6 % F1-score, outperforming individual CNN and Transformer baselines.
Recommended Citation
Elzayat, A. S., & Sayed, G. I. (2026, July). A Pretrained CNN–Transformer Gated Fusion Architecture for Ovarian Cancer Histological Subtype. In 2026 Intelligent Methods, Systems, and Applications (IMSA) (pp. 306-311). IEEE. https://doi.org/10.1109/IMSA70415.2026.11700256