INFORMATIVITY-AWARE WEIGHTED CNN ENSEMBLE FOR UTERINE FIBROID ULTRASOUND IMAGE CLASSIFICATION

Ultrasound-based assessment of uterine fibroids can be affected by speckle noise, low contrast, weak tissue boundaries, and variations in image quality. This study presents a two-stage computational approach that combines quantitative ultrasound image informativity assessment with a validation-based weighted ensemble of convolutional neural networks (CNNs). In the first stage, each image is characterized using contrast, entropy, texture, gradient magnitude, and an estimated noise level. The normalized features are combined into an integral informativity score, which is used to distinguish informative from non-informative images. In the second stage, five CNN architectures-VGG16, VGG19, ResNet18, EfficientNetB0, and EfficientNetB1-produce class-probability estimates. Their ensemble weights are calculated from validation performance using the mean of accuracy, area under the ROC curve (AUC), recall, specificity, and F1-score. The reported informativity-screening results increased from 87.88% to 95.24% accuracy, while the dataset was divided into 2,933 informative and 117 non-informative images. For classification, the proposed weighted ensemble achieved 93.88% accuracy, 0.9583 AUC, 95.39% recall, 89.36% specificity, and 90.06% F1-score. These results suggest that combining image-quality assessment with validation-informed ensemble weighting may improve the robustness of ultrasound image classification. Further evaluation on an independently collected dataset is required before clinical use.