Deep learning could play a key role in the workflow for renal ultrasound exams, potentially providing a preliminary diagnosis before the radiologist has had a chance to get to the study, according to study co-author Dr. Hersh Sagreiya of Stanford University in Palo Alto, CA.
"This is especially important as radiologist workflows increase and the demands from clinicians -- and especially the [emergency department] -- increase," Sagreiya noted to AuntMinnie.com.
However, ultrasound is a challenging modality for this type of analysis due to the significant heterogeneity in how these images are acquired, he said. In their project, the researchers sought to focus on the ability of deep learning to detect cystic lesions on ultrasound.
They trained a variety of deep-learning models using approximately 2,200 ultrasound exams, which had been produced on different scanners and labeled as either normal or abnormal and also whether a cystic lesion was present.
In testing on a separate set of studies that were labeled from the consensus of two out of three radiologists, the researchers found that the model that performed best employed a concept known as visual attention. This visual attention model enables the network to essentially "pay attention" to the subset of the images that are most informative, according to Sagreiya.
"We hope that this approach could aid in the triage of patients undergoing renal ultrasound examinations," Sagreiya said. "In the future, we may focus on other abnormalities, such as renal stones and hydronephrosis, which occurs when there is a blockage to the kidney. In addition, this technique could be applied to other types of ultrasound examinations."
Learn all about this research by sitting in on this presentation on Thursday.
