MP29-18: Development of an automated urinary stone detection and diagnosis system using artificial intelligence
Saturday, May 4, 2024 9:30 AM to 11:30 AM · 2 hr. (US/Central)
302B
Abstract
Information
Full Abstract and Figures
Author Block
Junghoon Lee, Sangjun Yoo, Min Chul Cho, Hyeon Jeong, Hwancheol Son, minsoo choo*, Dongjak-gu, Korea, Republic of
Introduction
we purposed to develop an AI algorithm that automatically detects urinary stones in CT images using an advanced deep learning architecture, and automatically calculates stone characteristic parameters such as volume or density that are essential for determining treatment methods. Furthermore, the performance of this novel program was compared with that of urologists within real-world emergency room scenarios.
Methods
The deep learning algorithm was trained using axial images of non-enhanced CT from patients who underwent stone surgery between August 2022 and July 2023. This dataset, comprising images with and without stones, was randomly partitioned into training, validation, and test sets. Two urologists and an AI specialist independently annotated stones in images using ‘Labelimg’ to create ground-truth data. The algorithm was trained using the YOLOv4 deep learning system. The AI model was validated externally using CT images of 100 consecutive patients who visited the emergency room with symptoms of suspected urinary stones.
Results
A deep learning-based AI model was developed using 39,433 CT images, of which 3,570 (9.1%) were positive samples. The model achieved a detection accuracy of 95%, with performance peaking at a 90% maximum average precision when using a 1:2 ratio of positive to negative samples. With a validation dataset of 5,736 CT images, including 482 positive samples, an AI model achieved a 95% detection accuracy. Only 13 (2.6%) stones were missed, primarily small stones under 2mm, or linear or non-oval stones. The model misclassified 178 (3.4%) negative samples as positive, primarily due to small high-attenuation artifacts and minor vascular calcifications. The AI model was further evaluated in a prospective study of 100 consecutive ER patients for suspected urolithiasis, diagnosing 87 on CT scan. The AI model achieved a high accuracy of 94% in reading CT scans. These missed cases were all small distal ureter stones or UVJ stones, which were not learned in the training dataset. The AI outperformed human specialists in speed, taking 13 seconds to review approximately 150 CT images versus urologists' 38.6 seconds, and 23 hours for formal reading. For calculating stone volume, it took an average of 77 seconds for three urologists to measure its width, length, and height. The AI model, on the other hand, could compute all stone parameters within 0.2 seconds.
Conclusions
We have implemented a novel automated detection system using a deep learning algorithm to assist clinicians in the diagnosis and characterization of urinary stones. The AI system demonstrated an accuracy of 94% in real clinical scenarios within emergency rooms, highlighting its potential to revolutionize the diagnosing speed in a general computing environment with consumer-grade GPUs.
Source Of Funding
Supported by grant No. '0320212070' for the SNUH Research Fund.
Author Block
Junghoon Lee, Sangjun Yoo, Min Chul Cho, Hyeon Jeong, Hwancheol Son, minsoo choo*, Dongjak-gu, Korea, Republic of
Introduction
we purposed to develop an AI algorithm that automatically detects urinary stones in CT images using an advanced deep learning architecture, and automatically calculates stone characteristic parameters such as volume or density that are essential for determining treatment methods. Furthermore, the performance of this novel program was compared with that of urologists within real-world emergency room scenarios.
Methods
The deep learning algorithm was trained using axial images of non-enhanced CT from patients who underwent stone surgery between August 2022 and July 2023. This dataset, comprising images with and without stones, was randomly partitioned into training, validation, and test sets. Two urologists and an AI specialist independently annotated stones in images using ‘Labelimg’ to create ground-truth data. The algorithm was trained using the YOLOv4 deep learning system. The AI model was validated externally using CT images of 100 consecutive patients who visited the emergency room with symptoms of suspected urinary stones.
Results
A deep learning-based AI model was developed using 39,433 CT images, of which 3,570 (9.1%) were positive samples. The model achieved a detection accuracy of 95%, with performance peaking at a 90% maximum average precision when using a 1:2 ratio of positive to negative samples. With a validation dataset of 5,736 CT images, including 482 positive samples, an AI model achieved a 95% detection accuracy. Only 13 (2.6%) stones were missed, primarily small stones under 2mm, or linear or non-oval stones. The model misclassified 178 (3.4%) negative samples as positive, primarily due to small high-attenuation artifacts and minor vascular calcifications. The AI model was further evaluated in a prospective study of 100 consecutive ER patients for suspected urolithiasis, diagnosing 87 on CT scan. The AI model achieved a high accuracy of 94% in reading CT scans. These missed cases were all small distal ureter stones or UVJ stones, which were not learned in the training dataset. The AI outperformed human specialists in speed, taking 13 seconds to review approximately 150 CT images versus urologists' 38.6 seconds, and 23 hours for formal reading. For calculating stone volume, it took an average of 77 seconds for three urologists to measure its width, length, and height. The AI model, on the other hand, could compute all stone parameters within 0.2 seconds.
Conclusions
We have implemented a novel automated detection system using a deep learning algorithm to assist clinicians in the diagnosis and characterization of urinary stones. The AI system demonstrated an accuracy of 94% in real clinical scenarios within emergency rooms, highlighting its potential to revolutionize the diagnosing speed in a general computing environment with consumer-grade GPUs.
Source Of Funding
Supported by grant No. '0320212070' for the SNUH Research Fund.