TY - JOUR

T1 - Deep learning for detection of clinical operations in robot-assisted percutaneous renal access

AU - Ibragimov, Bulat

AU - Zhen, Janet

AU - Ayvali, Elif

N1 - Publisher Copyright: Author

PY - 2023

Y1 - 2023

N2 - Percutaneous nephrolithotomy (PCNL) is the current standard of care for patients with a total renal stone burden > 20 mm. Gaining access to the kidney is a crucial step as the position of the percutaneous tract can affect the ability to manipulate a nephroscope during the procedure. However, gaining percutaneous access using fluoroscopic guidance has a challenging learning curve with only a minority of urologists can successfully establish the access. In addition to difficult access, the PCNL carries a risk of bleeding and need for blood transfusion. Robotic assistance may be a key towards accurate and reliable access. Beyond assisting with renal access, a robotic platform can record data of importance related to the user’s activities via sensor-equipped instruments. The analysis of these activities is crucial for understanding what constitutes a successful and safe procedure. In this paper, we harness the powers of machine learning to automatically analyze physicians’ activities during robotic-assisted renal access using the Monarch® Platform, Urology. A machine learning framework based on a combination of a 1-dimensional U-net and random forests was developed to find consistent patterns in the sensor data characteristic of needle insertions. This framework was used to retrospectively analyze data previously obtained from 248 percutaneous renal access procedures. These procedures were performed on 18 human cadaveric models by 17 practicing urologists and one urologist proxy. The framework automatically recognized 94% of all first needle insertions in each procedure and labeled them with the accuracy of 0.81 measured in terms of the Dice coefficient. The recognition accuracy for secondary insertions was 66%. The automatically detected needle insertions were used to calculate clinical metrics such as tract length, anterior-posterior and cranial-caudal angles of the insertion site, as well as user skills such as trajectory deviation and targeting accuracy.

AB - Percutaneous nephrolithotomy (PCNL) is the current standard of care for patients with a total renal stone burden > 20 mm. Gaining access to the kidney is a crucial step as the position of the percutaneous tract can affect the ability to manipulate a nephroscope during the procedure. However, gaining percutaneous access using fluoroscopic guidance has a challenging learning curve with only a minority of urologists can successfully establish the access. In addition to difficult access, the PCNL carries a risk of bleeding and need for blood transfusion. Robotic assistance may be a key towards accurate and reliable access. Beyond assisting with renal access, a robotic platform can record data of importance related to the user’s activities via sensor-equipped instruments. The analysis of these activities is crucial for understanding what constitutes a successful and safe procedure. In this paper, we harness the powers of machine learning to automatically analyze physicians’ activities during robotic-assisted renal access using the Monarch® Platform, Urology. A machine learning framework based on a combination of a 1-dimensional U-net and random forests was developed to find consistent patterns in the sensor data characteristic of needle insertions. This framework was used to retrospectively analyze data previously obtained from 248 percutaneous renal access procedures. These procedures were performed on 18 human cadaveric models by 17 practicing urologists and one urologist proxy. The framework automatically recognized 94% of all first needle insertions in each procedure and labeled them with the accuracy of 0.81 measured in terms of the Dice coefficient. The recognition accuracy for secondary insertions was 66%. The automatically detected needle insertions were used to calculate clinical metrics such as tract length, anterior-posterior and cranial-caudal angles of the insertion site, as well as user skills such as trajectory deviation and targeting accuracy.

KW - Human factors

KW - human performance analysis

KW - Kidney

KW - Kidney stones

KW - Machine learning

KW - machine learning

KW - Medical robotics

KW - Needles

KW - Percutaneous nephrolithotomy

KW - Performance evaluation

KW - Phantoms

KW - Robot sensing systems

KW - robotic surgery

KW - Robots

KW - Trajectory

UR - http://www.scopus.com/inward/record.url?scp=85168260034&partnerID=8YFLogxK

U2 - 10.1109/ACCESS.2023.3305246

DO - 10.1109/ACCESS.2023.3305246

M3 - Journal article

AN - SCOPUS:85168260034

VL - 11

SP - 90358

EP - 90366

JO - IEEE Access

JF - IEEE Access

SN - 2169-3536

ER -