A Deep Learning Model to automate bladder volume using Clarius Handheld Ultrasound
Location
Suwanee, GA
Start Date
7-5-2024 1:00 PM
End Date
7-5-2024 4:00 PM
Description
Introduction
Lower urinary tract symptoms (LUTS), including voiding and storage symptoms, have been universally recognized since the standardization of the urinary tract function terminology by the International Continence Society in 2002. Assessing bladder volume can determine the type of LUTS. A common manual method of determining volume from 2D Ultrasound (US) images uses the prolate ellipsoid method, requiring sagittal and transverse plane measurements with on-screen calipers.. The process is repetitive, time-consuming, and requires training to acquire the requisite skills. Artificial intelligence (AI) using machine learning algorithms can provide accurate measurement by identifying the bladder region and eliminating artifacts or noise from the image. This can improve image quality, reduce operator variability, and improve real-time diagnostic accuracy.
Objectives
The primary objective of this non-inferior prospective study was to compare bladder volume AI (BLAI) measurement to human measurement. A secondary objective determined if BLAI correctly identified the imaging plane.
Methods
The study was approved by PCOM's institutional review board (IRB) and included non-pregnant individuals over 18 years who provided informed consent. Fifty-eight subjects (18 Males: 40 Females) with a mean age of 31 years and mean BMI of 26.88 kg/m2 were recruited. Three physical therapists (PTs) performed all scans using Clarius C3HD (curvilinear) and PAHD (phased array) scanners. Images were uploaded to the Clarius Cloud and blindly reviewed by the PTs, who independently and manually measured bladder volume via the prolate ellipsoid method, indicated the bladder view (sagittal or transverse), and traced the bladder wall for each image. The absolute difference between reviewer pairs was calculated and compared to the absolute difference between BLAI. Mean reviewer measurements using a t-test and an equivalence margin of 25% (i.e., the mean difference between differences should be no greater than 25% of the measured bladder difference). Inter-rater reliability was determined via the intraclass correlation coefficients, with the accuracy of image segmentation measurements determined with the average Dice similarity coefficient and Jaccard similarity index.
Results
BLAI volume measurement was non-inferior (p<.001), with a mean difference between the human (PTs) and BLAI of -0.0228ml (95% CI -0.074, 0.028). The IRR between BLAI and individual human measurements indicated strong overall agreement. The ICC between reviewer 1 vs reviewer 2 was 0.98 (95% CI 0.96, 0.99), reviewer 1 vs reviewer 3 was 0.98 (95% CI 0.97,0.99), reviewer 2 vs. reviewer 3 was 0.98 (95% CI 0.97, 0.99), and BLAI to the reviewer mean was 0.91 (95% CI 0.85,0.95), indicating excellent reliability. Dice coefficients for reviewers 1, 2, and 3 compared to BLAI were 0.93 (95% CI 0.92, 0.93), 0.92 (95% CI 0.91, 0.93), and 0.93 (95% CI 0.92, 0.93), respectively. The Jaccard similarity index for reviewers 1, 2 & 3 was 0.87 (95% CI 0.85, 0.99), 0.86 (95% CI 0.84, 0.87), and 0.86 (95% CI 0.85, 0.88) respectively, further supporting measurement accuracy.
Conclusion
Results confirm high levels of agreement and consistency across all bladder volume measurements (BLAI and human). The study's secondary objective for BLAI to correctly identify transverse and sagittal views was also successfully met.
Embargo Period
5-23-2024
A Deep Learning Model to automate bladder volume using Clarius Handheld Ultrasound
Suwanee, GA
Introduction
Lower urinary tract symptoms (LUTS), including voiding and storage symptoms, have been universally recognized since the standardization of the urinary tract function terminology by the International Continence Society in 2002. Assessing bladder volume can determine the type of LUTS. A common manual method of determining volume from 2D Ultrasound (US) images uses the prolate ellipsoid method, requiring sagittal and transverse plane measurements with on-screen calipers.. The process is repetitive, time-consuming, and requires training to acquire the requisite skills. Artificial intelligence (AI) using machine learning algorithms can provide accurate measurement by identifying the bladder region and eliminating artifacts or noise from the image. This can improve image quality, reduce operator variability, and improve real-time diagnostic accuracy.
Objectives
The primary objective of this non-inferior prospective study was to compare bladder volume AI (BLAI) measurement to human measurement. A secondary objective determined if BLAI correctly identified the imaging plane.
Methods
The study was approved by PCOM's institutional review board (IRB) and included non-pregnant individuals over 18 years who provided informed consent. Fifty-eight subjects (18 Males: 40 Females) with a mean age of 31 years and mean BMI of 26.88 kg/m2 were recruited. Three physical therapists (PTs) performed all scans using Clarius C3HD (curvilinear) and PAHD (phased array) scanners. Images were uploaded to the Clarius Cloud and blindly reviewed by the PTs, who independently and manually measured bladder volume via the prolate ellipsoid method, indicated the bladder view (sagittal or transverse), and traced the bladder wall for each image. The absolute difference between reviewer pairs was calculated and compared to the absolute difference between BLAI. Mean reviewer measurements using a t-test and an equivalence margin of 25% (i.e., the mean difference between differences should be no greater than 25% of the measured bladder difference). Inter-rater reliability was determined via the intraclass correlation coefficients, with the accuracy of image segmentation measurements determined with the average Dice similarity coefficient and Jaccard similarity index.
Results
BLAI volume measurement was non-inferior (p<.001), with a mean difference between the human (PTs) and BLAI of -0.0228ml (95% CI -0.074, 0.028). The IRR between BLAI and individual human measurements indicated strong overall agreement. The ICC between reviewer 1 vs reviewer 2 was 0.98 (95% CI 0.96, 0.99), reviewer 1 vs reviewer 3 was 0.98 (95% CI 0.97,0.99), reviewer 2 vs. reviewer 3 was 0.98 (95% CI 0.97, 0.99), and BLAI to the reviewer mean was 0.91 (95% CI 0.85,0.95), indicating excellent reliability. Dice coefficients for reviewers 1, 2, and 3 compared to BLAI were 0.93 (95% CI 0.92, 0.93), 0.92 (95% CI 0.91, 0.93), and 0.93 (95% CI 0.92, 0.93), respectively. The Jaccard similarity index for reviewers 1, 2 & 3 was 0.87 (95% CI 0.85, 0.99), 0.86 (95% CI 0.84, 0.87), and 0.86 (95% CI 0.85, 0.88) respectively, further supporting measurement accuracy.
Conclusion
Results confirm high levels of agreement and consistency across all bladder volume measurements (BLAI and human). The study's secondary objective for BLAI to correctly identify transverse and sagittal views was also successfully met.