Weighted Fusion Architecture for Improved Ovarian Tumor Segmentation from Ultrasound Images

Ovarian tumor segmentation is a challenging task that requires high accuracy. This study presents a novel method for improving binary segmentation in ovarian ultrasound imaging by using a weighted fusion approach. This strategy utilizes a grid search algorithm to optimize the weights for fusing segmented maps from different models, achieving higher Precision and Recall. We investigated 3 backbone architectures, including Inceptionv3, DenseNet121, and MobileNet within the UNet framework - renowned for its effectiveness in ultrasound imaging, to generate a fused heatmap. In addition to this, MobileNet was then adopted as the backbone across two other different segmentation models: FPN and PSPNet due to its efficiency and high performance, followed by a similar weighted fusion approach together with UNet for the final predicted mask. Our findings show significant improvements in all key metrics, including Precision, Recall, IoU, and F1 Score, over individual methods.