Optimized EfficientNet for Detection of Endometrial Cancer Using Histopathological Images

The diagnosis of endometrial cancer is highly dependent on histopathological analysis, which is often time-consuming and subjective. This study introduces an optimized EfficientNet-B7.5DeltaR model for the automatic detection of stromal tissue, blood vessels, and neutrophils in tissue patches. The novelty of this work lies in the extension of compound scaling with a new ????-regularization factor and a widthwise convolution layer, enhancing generalization and reducing overfitting compared to the baseline EfficientNet. The proposed model was trained on 91 H&E-stained whole-slide images of endometrial carcinoma Pipelle biopsies from the Radboud University Medical Center, The Netherlands, and evaluated using standard classification and regression metrics. Results demonstrate a 6% improvement in validation accuracy, a 12% gain in R2 score, and a 5% increase in F1-score over the baseline, confirming the improved generalization of the model for histopathological cancer analysis.