Designing a Synthetic Ultrasound-Based Machine Learning Framework to Differentiate Renal Calculi From Ultrasound Artifacts and Evaluating Domain Transfer: Year 2

Overview

Kidney stones are solid mineral deposits that cause severe pain or harm to the urinary tract if not diagnosed and treated early. However, medical machine learning models require large labeled clinical datasets that are difficult to obtain. The goal of this project was to develop a machine learning framework that distinguished kidney stones from ultrasound artifacts using synthetic ultrasound images. This study investigated whether physics-based ultrasound simulation could generate data sufficient to train a convolutional neural network (CNN) with validation accuracy greater than 75%. Using Field II simulation software, 400 B-mode images were synthesized using controlled parameters. A ResNet-18 CNN was trained on images of stone-only and stone with reverberation artifacts. The model achieved 100% accuracy on the synthetic validation dataset, exceeding the target. To test real-world transfer, the model was evaluated on 1884 real ultrasound images (1001 stone, 883 no-stone). The model assigned higher probabilities to real stone-present images than to no-stone images (mean scores: 0.257 vs. 0.168). Statistical testing confirmed this difference was highly significant (KS p < 0.001, Mann-Whitney p < 0 .001). UMAP dimensionality reduction was used to map similarity between the two classes. A clear separation between clusters suggested that the learned features captured meaningful differences between real stone and no-stone images. These findings demonstrate the potential of synthetic images as an alternative training source for medical machine learning models when clinical data is limited. Future work will improve synthetic data realism using Generative Adversarial Networks (GANs) and additional simulated artifacts.

Competition history

ISEF 2026 Robotics and Intelligent Machines · Entry ROBO009

Resources

ISEF abstract page

Related projects

ISEF · 2021

A Novel Hybrid Deep Learning Model to Predict Acute Kidney Injury Using Patient Record Data and Ultrasound Kidney Images

ISEF · 2023

NephroNet: A Novel Program for Identifying Renal Cell Carcinoma and Generating Synthetic Training Images With Convolutional Neural Networks and Diffusion Models

ISEF · 2022

Detection of Benign and Malignant Lung Nodules in 3D Volumes Generated From Thoracic Computed Tomography Scans Leveraging Artificial Intelligence, Year 2

ISEF · 2019

Segmenting CT Slices: Optimizing Lesion Detection through Mask Region-based Convolutional Neural Networks

Closest projects by meaning, across every fair and year in the corpus.