DEEP CNN ENSEMBLE MODELS FOR ACCURATE OVARIAN CANCER DETECTION

Abstract

Ovarian cancer is still among the most lethal gynaecologic malignancies because it is detected too late and has blurred early manifestations. An early diagnosis is crucial to enhance the cure rates and survival rate. In this paper, we intend to propose a deep convolutional neural network ensemble framework based on transfer learning for the classification of ovarian tumours into three categories, namely Normal, Benign, and Malignant, with high accuracy. This ensemble combines three optimized models: a fine-tuned ResNet50, a VGG16 coupled with an FCNN designed in this work, and a fine-tuned EfficientNet-B0. All three models have been pre-trained on ImageNet and were adapted to the ovarian ultrasound domain by selective fine-tuning of layers and domain-specific optimization techniques.

in order to handle class imbalance and improve diagnostic reliability, class weights and Focal Loss were used (for ResNet50), along with extensive TTA to improve model generalization. Adaptive optimizers, early stopping, and checkpoint mechanisms were employed to ensure the convergence of the networks in a stable manner. Experimental evaluation shows that the proposed ensemble significantly improves the accuracy, sensitivity and specificity of classification as compared to the any standalone model, thereby offering the great potential in serving as an intelligent and reliable decision support system for early ovarian cancer screening. The study emphasizes how combining the transfer learning techniques with ensemble learning models can significantly enhance the performance of AI driven medical diagnostic systems, making them more accurate, reliable, and clinically useful for the real-world diagnosis healthcare applications.