Shadow Detection for Ultrasound Images Using Unlabeled Data and Synthetic Shadows

TL;DR

This paper introduces a novel auto-encoding based method for detecting shadows in ultrasound images, leveraging synthetic shadows and unlabeled data to improve diagnostic accuracy.

Contribution

It presents a new shadow detection approach that uses synthetic shadows and an auto-encoding structure, requiring only unlabeled data for training.

Findings

Effective shadow detection demonstrated on fetal heart ultrasound images.

Method outperforms existing techniques in shadow identification accuracy.

Utilizes synthetic shadows to guide the network without labeled data.

Abstract

Medical ultrasound is widely used technique for diagnosing internal organs. As common artifacts, shadows often appear in ultrasound images. Detecting such shadows is curious because they prevent accurate diagnosis. In this paper, we propose a novel shadow detection method based on auto-encoding structure. It once separates an input image into a shadow image and a content image using two decoders and combines them to reconstruct the input. To lead the network into separating the input, we inject synthetic shadows into the input and make the network to predict them as the shadow image. Since we know the rough shape of shadows as basic domain knowledge, we can generate plausible shadows. These processes are achieved by using only unlabeled data. Experiments on ultrasound images for fetal heart diagnosis shows the effectiveness of the method.