Optimasi Segmentasi Kepala Janin Berbasis U-Net Melalui Preprocessing Citra USG

Abstract

Fetal Head Circumference (HC) measurement using Ultrasound (USG) imagery is a crucial biometric parameter for estimating gestational age and monitoring fetal growth rate. However, automated interpretation is often hindered by speckle noise, low contrast, and blurred object boundaries inherent in USG images. This study aims to optimize the performance of a U-Net architecture with backbone ResNet-34 for fetal head segmentation through image preprocessing and data augmentation techniques. The proposed method integrates Anisotropic Diffusion for noise reduction and CLAHE (Contrast Limited Adaptive Histogram Equalization) to enhance boundary features, alongside geometric augmentations (rotation, flip) and median blur. The model was trained on 799 training images and validation with 80:20 ratio and 200 test images from a public dataset. Results indicate that the proposed preprocessing significantly improves segmentation performance compared to the baseline. The Intersection over Union (IoU) score increased from 0.9440 to 0.9526, and the Dice Similarity Coefficient (DSC) get 0.9757. Although preprocessing visually intensified certain artifacts, it effectively enhanced feature distinctiveness for the model. Based on the segmentation output, biometric estimation was conducted using ellipse fitting. This study concludes that U-Net optimized with Anisotropic Diffusion and CLAHE preprocessing shows significant potential as an assistive tool for medical professionals, enabling faster biometric measurement while maintaining the necessity for clinical verification.