Comparison of Two Augmentation Methods in Improving Detection Accuracy of Hemarthrosis

With the increase of computing power, machine learning models in medical imaging have been introduced to help in rending medical diagnosis and inspection, like hemophilia, a rare disorder in which blood cannot clot normally. Often, one of the bottlenecks of detecting hemophilia is the lack of data available to train the algorithm to increase the accuracy. As a possible solution, this research investigated whether introducing augmented data by data synthesis or traditional augmentation techniques can improve model accuracy, helping to diagnose the diseases. To tackle this research, features of ultrasound images were extracted by the pre-trained VGG-16, and similarities were compared by cosine similarity measure based on extracted features in different distributions among real images, synthetic images, and augmentation images (Real vs. Real, Syn vs. Syn, Real vs. Different Batches of Syn, Real vs. Augmentation Techniques). Model testing performance was investigated using EffientNet-B4 to recognize "blood" images with two augmentation methods. In addition, a gradient-weighted class activation mapping (Grad-CAM) visualization was used to interpret the unexpected results like loss of accuracy. Synthetic and real images do not show high similarity, with a mean similarity score of 0.4737. Synthetic batch 1 dataset and images by horizontal flip are more similar to the original images. Classic augmentation techniques and data synthesis can improve model accuracy, and data by traditional augmentation techniques have a better performance than synthetic data. In addition, the Grad-CAM heatmap figured out the loss of accuracy is due to a shift in the domain. Overall, this research found that two augmentation methods, data synthesis and traditional augmentation techniques, both can improve accuracy to a certain extent to help to diagnose rare diseases.

NeRF-US: Removing Ultrasound Imaging Artifacts from Neural Radiance Fields in the Wild

Current methods for performing 3D reconstruction and novel view synthesis (NVS) in ultrasound imaging data often face severe artifacts when training NeRF-based approaches. The artifacts produced by current approaches differ from NeRF floaters in general scenes because of the unique nature of ultrasound capture. Furthermore, existing models fail to produce reasonable 3D reconstructions when ultrasound data is captured or obtained casually in uncontrolled environments, which is common in clinical settings. Consequently, existing reconstruction and NVS methods struggle to handle ultrasound motion, fail to capture intricate details, and cannot model transparent and reflective surfaces. In this work, we introduced NeRF-US, which incorporates 3D-geometry guidance for border probability and scattering density into NeRF training, while also utilizing ultrasound-specific rendering over traditional volume rendering. These 3D priors are learned through a diffusion model. Through experiments conducted on our new "Ultrasound in the Wild" dataset, we observed accurate, clinically plausible, artifact-free reconstructions.