Novel noninvasive assessment of upper urinary tract urine flow dynamics: a deep learning-driven reconstruction model combined with CFD simulation

Abstract

Purpose

Recent advances in deep learning have facilitated the exploration of mesh reconstruction techniques; however, these approaches have not yet been applied to mesh reconstruction of hydronephrosis lesions in pediatric patients. Three surgical strategies are currently available for robot-assisted pyeloplasty in children, but there remains no consensus on which yields superior clinical benefits.

Materials and methods

A total of 36 patients received robot-assisted pyeloplasty were divided into three groups based on different surgical procedures: large anastomosis group (LAG), middle anastomosis group (MAG) and small anastomosis group (SAG).A deep learning network was utilized to reconstruct upper urinary tract models from patients’ magnetic resonance urography (MRU) images, followed by computational fluid dynamics (CFD) simulations to measure urine flow dynamics parameters in the reconstructed models. Statistical comparisons of parameters were performed among the three groups.

Results

Statistically significant differences were observed among the three groups in terms of improvement in differential renal function (DRF), reduction in postoperative anteroposterior diameter (APD) of the renal pelvis, postoperative pressure reduction, pressure gradient between the upper and lower planes, changes in urinary flow velocity, and alterations in wall eddy viscosity. DRF was significantly improved in both the MAG and LAG, whereas it was significantly decreased in the SAG. The MAG exhibited the most remarkable reduction in the anteroposterior diameter of the renal pelvis. Postoperative pressure reduction was more pronounced in the MAG when compared with the LAG and SAG. The median postoperative pressure gradient between the upper and lower planes was the highest in the LAG. Postoperative urinary flow velocity increased significantly in the LAG, only slightly in the MAG, but decreased in the SAG. Postoperative wall eddy viscosity increased significantly in the LAG, whereas it decreased in both the MAG and SAG.

Conclusions

Deep learning-based upper urinary tract model reconstruction combined with CFD simulation constitutes a novel noninvasive method for measuring upper urinary tract urine flow dynamics parameters. Among the three robot-assisted pyeloplasty techniques, the middle anastomosis approach yields the optimal postoperative recovery.

Introduction

Ureteropelvic junction obstruction (UPJO), defined as functional or anatomical impairment of urine flow from the renal pelvis to the proximal ureter, is the most common perinatal cause of hydronephrosis. Congenital UPJO affects approximately 1 in 1000 to 2000 live births, with a male-to-female incidence ratio of 2:1 [1].The gold standard treatment for UPJO is dismembered pyeloplasty, as described by Anderson and Hynes [2].

Traditionally, this surgery is performed via open or laparoscopic approaches. The da Vinci Robotic Surgical System represents the most advanced surgical platform currently available, marking a milestone in the integration of artificial intelligence and medical technology. Its high precision and flexibility are unparalleled by laparoscopic equipment. As early as 2004, Kant et al. noted that pediatric surgery is a highly promising field for robotic systems, particularly for complex reconstructive procedures in children with small body cavities [3]. Although robot-assisted pyeloplasty has improved surgical precision, individualized surgical planning remains lacking.

The fundamental principle of robot-assisted Anderson–Hynes pyeloplasty involves resecting the obstructed junction tissue and reanastomosing the renal pelvis and ureter. Initially, it was believed that surgery should not only remove the obstructed tissue but also excise part of the dilated renal pelvis, as dilation may lead to dysperistalsis and impair postoperative recovery. Consequently, three surgical viewpoints have emerged: (1) resect most of the dilated pelvis to within 1 cm of the renal parenchyma; (2) moderately resect the pelvis to 2–3 cm from the renal parenchyma, as excessive resection increases the risk of anastomotic fistula and stenosis, while residual dilation does not appear to affect urinary flow recovery; and (3) resect only the obstructed junction without reducing the dilated pelvis. These three approaches result in anastomoses of varying sizes (Fig. 1), and current surgical decision-making primarily relies on physician preference, with no consensus on which method optimizes postoperative recovery.

Fig. 1.

Schematic diagram of the three surgical approaches: a large anastomosis approach; b middle anastomosis approach; c small anastomosis appro

Classically, postoperative recovery assessment has relied on sonography, magnetic resonance urography (MRU), and nuclear renography; however, none of these modalities can provide direct urine flow dynamics parameters. In the present study, we propose a novel approach combining deep learning-based reconstruction models and CFD simulations to simulate pelvic and upper ureteral urine flow. This noninvasive method enables measurement of urine flow dynamics parameters (e.g., pressure and velocity) throughout the upper urinary tract, avoiding invasive procedures and improving the accuracy of recovery assessment.

Materials and methods

Inclusion and exclusion criteria

Inclusion criteria: The hydronephrosis was SFU Ⅲ ~ Ⅳ grade pre-operation with deteriorating renal function, abdominal pain or urinary tract infections. In addition, the etiology was ureteropelvic junction obstruction confirmed by postoperative pathology at the same time.

Exclusion criteria: The hydronephrosis was secondary to repetitive kidney, vesicoureteral reflux, megalouteral, neurogenic bladder and so on.

Patients and design

A total of 36 patients diagnosed with unilateral UPJO were received robot-assisted Anderson–Hynes pyeloplasty at our institution between October 2020 and January 2023.All procedures were performed by the same surgeon and the da Vinci Xi Surgical Robot (Da Vinci, Mountain View, CA, USA) was used. They were divided into three groups according to different surgical methods. The small anastomosis group(SAG), where only the stenotic segment was resected without renal pelvis reduction; the middle anastomosis group(MAG), where the stenotic segment was resected with 1 cm upward reduction of the renal pelvis; and the large anastomosis group(LAG), where the stenotic segment was resected with approximately 2 cm upward reduction of the renal pelvis.

The choice of surgical method was based mainly on the parent’s wishes. The patients’ basic information and medical record data of sonography, MRU and ⁹⁹ᵀᶜᵐ-DMSA renal static scintigraphy before and after operation were collected.

MRU

We used non-contrast-enhanced magnetic resonance urography.On the day of the examination oral hydration (250–500 ml of water or juice) is provided to the child in the waiting room 30 min before the examination. The administration of intravenous fluids for hydration is considered unnecessary [4].All children younger than 5 years required sedation with oral chloral hydrate according to the usual protocol established by the anesthetist. Sedated children were monitored through their breathing and heart rates. Toilet-trained children are requested to void their bladder before entering the MRI room as this avoids discomfort and premature interruption of the study and a full bladder is known to hamper upper urinary tract drainage. All patients were prepared without intravenous furosemide. All the MR examinations were performed in a supine position. All scans were performed using a 1.5 T MAGNETOM Area magnetic resonance machine (Siemens, Germany).MR sequences included single excited fast spin-echo (SSFSE) or half-Fourier acquisition single-shot turbo spin-echo (HASTE) in coronal position to cover the entire kidney and ureteral area, axial T1WI and fat-saturated T2-weighted sequence to cover the entire kidney, and heavily T2-weighted imaging (two-dimensional) in coronal position covering the entire urinary system. The 1.5 T pediatric non-contrast MRU had a total examination time of 12–20 min: coronal SSFSE/HASTE (entire kidneys and ureters) 8–15 s, axial T1WI (entire kidneys) 1.0–1.5 min, axial fat-saturated T2WI (entire kidneys) 1.2–2.0 min, coronal 2D heavily T2WI (entire urinary system) 1.5–2.5 min, with core sequences totaling 8–14 min. Cooperative children ≥ 3 years took 10–15 min, while restless children 1–3 years took 12–20 min with possible short re-scans.

Deep learning-based reconstruction of the upper urinary tract model

Upper urinary tract segmentation

The nnU-Net deep learning framework was utilized (Fig. 2), an explicit method integrating convolutional neural networks (CNNs) and graph convolutional networks (GCNs) for surface mesh reconstruction of hydronephrosis lesions from MRU images. nnU-Net is a U-shaped architecture with proven efficacy in various medical image segmentation tasks [5].

Fig. 2.

Deep learning-based upper urinary tract model reconstruction. a Data preprocessing: cropping and contrast enhancement were performed on raw MRU data. b Network framework: the nnU-Net architecture consists of an encoder, decoder, and skip connections. c Three-dimensional (3D) reconstruction: segmentation masks were converted into usable 3D mesh models via smoothing and meshing techniques

For upper urinary tract segmentation:

1. Data preprocessing: MRU data underwent cropping and contrast enhancement to improve segmentation accuracy;

2. Model training: Additional MRU data from 25 pediatric UPJO patients were collected, and each MRU image was independently annotated by two experienced radiologists to label the renal pelvis and calyces. Training included 500 iterations, and fivefold cross-validation was used to optimize model performance and reduce overfitting;

3. Segmentation implementation: The trained nnU-Net model was applied to unlabeled MRU data to automatically generate segmentation masks for the renal pelvis and calyces.

Mesh reconstruction

Segmentation results were further processed to generate CFD-compatible upper urinary tract mesh models (Fig. 2):

1. Conversion of segmentation labels into a 3D upper urinary tract model;

2. Model verification to ensure geometric integrity and eliminate structural defects;

3. Mesh smoothing to enhance surface quality and reduce simulation errors;

4. Surface reconstruction to repair potential topological defects;

5. Grid generation to create the final structured mesh structure.

Urine flow dynamics

CFD simulation

Urine was modeled as an incompressible Newtonian fluid, mainly composed of water, urea, electrolytes, and metabolic wastes, with continuous flow during urine formation. Physical properties included density (1000–1030 kg/m³), dynamic viscosity (0.6–2.5 mPa·s), specific heat capacity (≈4190 J/kg·K), thermal conductivity (≈0.6 W/m·K), and pH (4.5–8.0) [6]. Inlet urine velocity was assumed to be perpendicular to the plane and uniformly distributed. The walls of the renal pelvis, calyces, and ureters were considered as continuous soft tissues (density: 1000–1100 kg/m³; dynamic viscosity: 0.002–0.004 mPa·s; specific heat capacity: 3500–4200 J/kg·K; thermal conductivity: 0.4–0.6 W/m·K; thermal expansion coefficient: ≈0.0002 1/K) [7] that support and guide urine flow.

Fluid inlet regions were defined at the renal calyces, and the outlet at the ureter. Due to interpatient variability in calyx number, the number of inlet regions varied among patients. Thus, direct comparisons were restricted to preoperative and postoperative data within the same patient. CFD simulations used double-precision calculations to accurately capture fluid dynamics and urine mechanical properties [8]. The energy equation and k-epsilon turbulence model (residual threshold: 10⁻⁴) were employed. Transient simulations were based on actual patient urodynamic data. To stabilize the simulation, the initial velocity was set to closely mimic the physiological urine production process, with an inlet velocity of 0.0001 m/s. Final results were obtained after 10 hybrid initialization calculations and 200 iterative operations (Fig. 3).

Fig. 3.

Upper urinary tract modeling for computational fluid dynamics (CFD) simulations

Urine flow analysis

Preoperative model analysis was used as an illustration (Fig. 4):

1. Three equidistant X–Y planes were extracted using the plane extractor tool in the CFD system: upper plane, middle plane, and lower plane (i.e., ureteropelvic junction plane);

2. Local cloud maps of pressure, velocity, and eddy viscosity were generated for the three planes;

3. Data from the cloud maps were extracted, and mean values were calculated for quantitative analysis;

4. Wall shear pressure and wall eddy viscosity maps were generated, with mean values calculated using the CFD post-processing module.

Fig. 4.

Urine flow dynamics parameters analysis of the preoperative model: a three equidistant X–Y planes (upper plane, middle plane, and lower plane [i.e., ureteropelvic junction plane]) were extracted using the plane extractor tool in the fluent system; b–d local cloud maps of pressure, velocity, and eddy viscosity; e, f wall shear pressure and wall eddy viscosity maps

Postoperative model analysis followed the same protocol. Direct comparisons were made between preoperative and postoperative data within the same patient.

Parameter definitions

• P: Percentage reduction in lower plane pressure postoperatively compared with preoperatively;

• ΔP: Pressure gradient between the upper and lower planes;

• V: Percentage change in lower plane velocity postoperatively compared with preoperatively;

• VISJ: Percentage change in lower plane eddy viscosity postoperatively compared with preoperatively;

• VISZ: Percentage change in wall eddy viscosity postoperatively compared with preoperatively;

• WSS: Percentage change in wall shear pressure postoperatively compared with preoperatively.

Statistical analysis

Statistical analyses were performed using SPSS 21.0 software (IBM Corp., Armonk, NY, USA). Categorical variables were presented as counts and percentages, and non-normally distributed continuous variables were presented as median (interquartile range [IQR]). Nonparametric tests were used for comparisons: Kruskal–Wallis H test for intergroup comparisons, and Mann–Whitney U test for pairwise comparisons. A two-tailed p value < 0.05 was considered statistically significant.

Results

Baseline and outcome data of each group

Large anastomosis group (LAG)

The median age was 19.00 (IQR 8.00–37.00) months, with a median follow-up duration of 15.50 (IQR 9.75–19.25) months. DRF improved slightly, with a median increase of 1.92% (IQR 0.45–3.70%). Pressure was reduced by a median of 65.59% (IQR 22.41–71.81%), with a median ΔP of 83.57 (IQR 51.23–202.08) Pa. Velocity increased by a median of 46.87% (IQR 20.97–129.45%). Wall shear pressure increased by a median of 0.48% (IQR − 36.14 to 80.11%), wall eddy viscosity increased by a median of 16.46% (IQR − 0.10 to 27.46%), and lower plane eddy viscosity decreased by a median of − 73.64% (IQR − 88.32 to − 30.70%) (Table 1).

Table 1.

Urine flow dynamics parameters of the large anastomosis group (LAG)

N	Age (M)	Sex (F/M)	side	SFU	DRF (%)	P (%)	ΔP (Pa)	V (%)	wss(%)	visz(%)	Visj (%)
1	12	M	R	Ⅳ	0.20	73.64	235.88	155.82	− 73.67	1.87	− 95.94
2	84	M	L	Ⅲ	1.35	8.67	106.57	− 2.61	20.30	− 11.76	− 34.14
3	2	M	L	Ⅳ	− 1.98	64.67	612.10	315.2	− 31.61	− 6.00	52.36
4	10	F	R	Ⅳ	3.49	3.50	47.44	26.52	77.26	33.73	− 81.40
5	2	M	R	Ⅳ	4.33	68.82	33.12	19.86	− 19.35	14.47	− 56.79
6	60	M	R	Ⅳ	4.40	26.50	52.49	58.13	202.22	26.54	− 85.78
7	1	F	L	Ⅳ	5.89	84.00	208.09	178.48	106.60	123.60	− 151.01
8	13	F	L	Ⅳ	− 1.67	10.12	150.35	65.23	70.23	20.33	− 66.5
9	36	F	R	Ⅳ	2.33	71.2	200.07	120.66	88.65	15.67	− 120.34
10	27	M	L	Ⅲ	3.15	30.8	60.56	35.61	− 34.67	− 9.88	− 20.36
11	40	M	L	Ⅲ	1.5	66.5	55.67	21.34	− 40.56	17.25	40.12
12	25	M	R	Ⅳ	0.53	80.56	40.55	15.45	− 57.66	30.23	− 80.78

Middle anastomosis group (MAG)

The median age was 84.00 (IQR 15.50–96.00) months, with a median follow-up duration of 14.50 (IQR 12.75–18.50) months. DRF improved significantly, with a median increase of 3.96% (IQR 1.99–12.68%). Pressure was reduced by a median of 90.42% (IQR 80.39–95.34%), with a median ΔP of 69.45 (IQR 45.28–131.01) Pa. Velocity increased slightly by a median of 7.21% (IQR − 32.08 to 36.29%). Wall shear pressure decreased by a median of − 42.74% (IQR − 60.89 to 12.07%), wall eddy viscosity decreased by a median of − 4.08% (IQR − 33.21 to 6.84%), and lower plane eddy viscosity decreased by a median of − 42.38% (IQR − 83.33 to − 15.22%) (Table 2).

Table 2.

Urine flow dynamics parameters of middle anastomosis group(MAG)

N	Age(M)	Sex (F/M)	side	SFU	DRF (%)	P(%)	ΔP (Pa)	v (%)	Wss (%)	Visz (%)	Visj (%)
1	120	M	L	Ⅲ	2.42	90.61	123.61	38.36	− 60.47	6.25	− 16.76
2	84	M	L	Ⅲ	0.12	63.68	55.05	92.71	12.03	13.59	− 95.25
3	7	M	L	Ⅲ	14.19	83.68	102.00	4.19	9.38	6.38	− 90.76
4	24	F	R	Ⅳ	0.72	94.38	26.47	32.82	12.19	14.49	− 13.05
5	84	M	L	Ⅳ	18.25	95.89	9.10	− 55.66	− 62.14	− 3.37	− 83.64
6	3	F	L	Ⅲ	4.46	96.19	72.20	− 47.16	− 84.68	− 52.2	− 38.86
7	108	M	L	Ⅲ	16.33	95.38	256.33	− 37.97	− 55.81	− 42.06	− 83.22
8	30	M	L	Ⅳ	3.45	90.23	153.2	35.6	− 40.28	− 30.26	− 45.89
9	10	F	R	Ⅳ	6.43	89.6	30.47	− 23.66	− 45.2	− 43.23	− 13.9
10	70	F	R	Ⅳ	12.18	95.32	50.22	10.23	15.66	− 15.2	− 67.89
11	36	F	R	Ⅳ	2.3	70.51	160.7	80.16	15.34	8.23	− 15.66
12	48	F	R	Ⅳ	1.06	63.87	66.69	− 30.12	− 67.87	− 4.78	− 10.18

Small anastomosis group (SAG)

The median age was 35.50 (IQR 24.00–67.50) months, with a median follow-up duration of 18.50 (IQR 15.50–26.75) months. DRF did not improve, with a median decrease of -1.10% (IQR − 4.60 to 1.93%). Pressure was reduced by a median of 59.25% (IQR 47.16–65.09%), with a median ΔP of 27.67 (IQR 16.83–59.86) Pa. Velocity decreased by a median of − 33.84% (IQR − 60.16 to − 28.46%). Wall shear pressure decreased by a median of − 32.70% (IQR − 42.34 to − 21.32%), wall eddy viscosity decreased by a median of − 15.72% (IQR − 22.63 to − 9.18%), and lower plane eddy viscosity decreased by a median of − 53.02% (IQR − 76.92 to − 43.25%) (Table 3).

Table 3.

Urine flow dynamics parameters of small anastomosis group(SAG)

N	Age(M)	Sex (F/M)	side	SFU	DRF (%)	P (%)	ΔP (Pa)	V (%)	Wss (%)	Visz (%)	Visj (%)
1	24	F	L	Ⅳ	− 14.92	30.60	79.55	− 71.6	− 18.01	− 1.69	− 86.96
2	24	M	L	Ⅳ	− 4.57	48.49	72.54	− 28.9	− 57.37	− 30.14	− 74.22
3	132	M	L	Ⅳ	− 3.13	56.30	0.23	− 35.54	− 60.97	− 7.10	− 27.37
4	72	M	L	Ⅲ	− 0.89	86.65	15.58	− 64.60	− 48.33	10.00	− 93.07
5	24	F	L	Ⅲ	1.90	73.50	10.73	− 64.49	− 30.75	− 42.36	− 50.30
6	2	M	L	Ⅲ	3.38	58.87	23.35	− 32.13	− 20.58	− 18.18	− 44.29
7	96	M	R	Ⅳ	4.20	62.94	59.63	− 36.65	− 35.28	− 19.18	− 55.66
8	56	M	R	Ⅳ	1.1	59.62	60.56	− 27.15	− 19.07	− 9.87	− 60.56
9	37	F	L	Ⅳ	2	43.17	57.26	− 10.6	− 34.65	− 13.25	− 36.9
10	66	M	L	Ⅲ	− 5.7	30.85	24.78	− 58.72	− 21.56	− 20.13	− 85
11	34	M	R	Ⅲ	− 1.3	70.56	30.56	− 30.58	− 40.34	− 35.67	− 40.12
12	25	M	R	Ⅲ	− 4.7	63.27	17.25	− 27.14	− 23.58	− 12.87	− 50.38

Intergroup comparisons

No significant differences were observed among the three groups in terms of age, follow-up duration, affected side, gender, preoperative SFU grade or pre-APD (all P > 0.05). However, significant differences existed in postoperative APD (post-APD) reduction, median diameter reduction fraction (DRF) improvement, median pressure reduction, median Δ P, median velocity change, and median wall eddy viscosity variation (all P < 0.05) (Table 4).

Table 4.

Intergroup comparison of urine flow dynamics parameters

	LAG	MAG	SAG	χ2	P
Sex (F:M)	4:8	6:6	3:9	1.69	0.43
Side (L:R)	6:6	7:5	8:4	0.67	0.71
SFU (Ⅲ:Ⅳ)	3:9	5:7	6:6	1.44	0.64
Age (M)	19.00(8.00.37.00)	42.00(20.50,84.00)	35.50(24.00,67.50)	3.96	0.14
Following-time(M)	15.50(9.75,19.25)	14.50(12.75,18.50)	18.50(15.50,26.75)	3.36	0.19
Pre-APD(mm)	37.07 ± 9.10	34.16 ± 8.42	35.35 ± 9.65	0.31	0.73
Post-APD(mm)	17.37 ± 6.64	10.37 ± 3.41	18.64 ± 12.51	3.36	0.047
DRF (%)	1.92(0.45,3.70)	3.96(1.99,12.68)	− 1.10(− 4.60,1.93)	18.50	< 0.001
P (%)	65.59(22.41,71.81)	90.42(80.39,95.34)	59.25(47.16,65.09)	12.67	0.002
ΔP (Pa)	83.57(51.23,202.08)	69.45(45.28,131.01)	27.67(16.83,59.86)	11.17	0.004
V (%)	46.87(20.97,129.45)	7.21(-32.08,36.29)	− 33.84(− 60.16, − 28.46)	12.17	0.002
Wss (%)	0.48(-36.14,80.11)	− 42.74(− 60.89,12.07)	− 32.70(− 42.34, − 21.32)	2.17	0.34
Visz (%)	16.46(-0.10,27.46)	− 4.08(− 33.21,6.84)	− 15.72(− 22.63, − 9.18)	12.16	0.002
Visj (%)	− 73.64(− 88.32, − 30.70)	− 42.38(− 83.33, − 15.22)	− 53.02(− 76.92, − 43.25)	1.17	0.56

Pairwise comparisons showed:

DRF improvement: The SAG differed significantly from the LAG (P = 0.04) and MAG (P = 0.003), but no significant difference was noted between the LAG and MAG (P = 0.08);

Post-APD: The MAG differed significantly from the LAG (P = 0.05) and SAG (P = 0.02), but no significant difference was noted between the LAG and SAG (P = 0.71).

Pressure reduction: Significant differences were observed between the LAG and MAG (P = 0.002), and between the MAG and SAG (P < 0.001), but not between the LAG and SAG (P = 0.95);

ΔP: The SAG differed significantly from the LAG (P = 0.01) and MAG (P < 0.001), but no significant difference was noted between the LAG and MAG (P = 0.45);

Velocity change: The MAG differed significantly from the LAG (P = 0.028) and SAG (P = 0.01), with a significant difference also observed between the LAG and SAG (P < 0.001);

Wall eddy viscosity variation: The LAG differed significantly from the MAG (P = 0.009) and SAG (P < 0.001), but no significant difference was noted between the MAG and SAG (P = 0.42).

In summary, DRF was significantly improved in both the MAG and LAG, but decreased in the SAG. The MAG exhibited the most remarkable reduction in the anteroposterior diameter of the renal pelvis. Pressure reduction was most pronounced in the MAG. ΔP was highest in the LAG. Velocity increased significantly in the LAG, slightly in the MAG, but decreased in the SAG. Wall eddy viscosity increased in the LAG but decreased in both the MAG and SAG.

Discussion

With the advancement of artificial intelligence technology, the da Vinci Robotic Surgical System was approved for clinical use by the FDA in 2000, ushering in a new era of surgical innovation. The key advantages of robotic technology include tremor filtration, motion scaling, and articulated endoscopic instruments (compared with traditional non-articulated ones), which are particularly crucial for complex surgical procedures. This has advanced the development of minimally invasive surgery and been successfully applied in multiple fields, including gastrointestinal and hepatobiliary surgery. In 2002, Dr. Craig Peters from Boston Children’s Hospital reported the first robot-assisted urologic surgery in a pediatric patient [9]., and in 2015, Lü Yiqing et al. performed the first robot-assisted Anderson–Hynes pyeloplasty in China [10]. After two decades of development, numerous studies have confirmed that robot-assisted pyeloplasty achieves higher precision than conventional laparoscopic surgery, even surpassing open surgery in delicacy [11]. Since 2020, our department has optimized and successfully implemented single-port-plus-one robot-assisted surgery in pediatric patients, publishing over 30 papers on this minimally invasive technique in peer-reviewed journals and presenting it at multiple national and international conferences [12, 13].In 2023, this technique was awarded the First Prize of the Fujian Medical Science and Technology Award.

Accurate assessment of obstruction severity and renal function is essential for optimizing surgical planning and predicting prognosis. The Whitaker test is a traditional invasive method for evaluating UPJO in clinical practice [14]^. This procedure involves placing a renal puncture fistula in advance, infusing sterile saline through the fistula at a rate of 10 mL/min, and measuring renal pelvic pressure (< 15 cm H₂O = no obstruction; > 22 cm H₂O = obstruction). However, this test is highly invasive, causes significant discomfort in pediatric patients, and is not recommended for routine clinical application. Sampaio pioneered the development of three-dimensional models of the renal pelvis and calyces using polyester resin corrosion, based on 146 renal units from 73 adult cadavers [15].Currently, some urologists have established computer models to investigate the urinary flow mechanism in adult calculous urinary tract obstruction, primarily using the renal pelvis and calyx morphological data reported by Sampaio [16, 17]. However, pediatric kidney morphology differs significantly from that of adults, rendering adult kidney models unsuitable for pediatric cases. Thus, there is an urgent need for precise and pediatric-specific modeling methods in this field.

Recent advances in deep learning have facilitated mesh reconstruction, particularly in cortical and white matter reconstruction [18, 19]. However, these methods have not been applied to mesh reconstruction of hydronephrosis lesions in pediatric patients. High-quality, accurate, and smooth models combined with CFD simulations offer a promising strategy for urodynamic evaluation, with demonstrated significant potential in disease diagnosis and analysis, especially in the cardiovascular and respiratory systems [20, 21]. In the urinary system, CFD is mainly used to study urinary flow dynamics and guide treatment strategies [22].and recent studies have explored its application in urinary stent flow analysis [23, 24]. Nevertheless, there is a significant research gap in pediatric urological modeling both domestically and internationally.

In the present study, we combined deep learning-based reconstruction models with CFD simulations to measure upper urinary tract urodynamic parameters. We found that lower plane pressure was significantly reduced postoperatively in patients who underwent successful pyeloplasty, with the most pronounced reduction in the MAG, followed by the LAG and SAG. The MAG exhibited the optimal median DRF and post-APD improvement postoperatively, followed by the LAG, while the SAG showed a decrease. These results indicate that DRF and post-APD improvement may be associated with pressure reduction, which is consistent with findings from previous studies. The median ΔP between the upper and lower planes was highest in the LAG, followed by the MAG and SAG. Velocity increased in the LAG and MAG but decreased in the SAG, suggesting a positive correlation between the pressure gradient and lower plane velocity. An appropriate pressure gradient between the upper and lower planes is necessary to enhance urinary flow velocity. Median wall shear pressure and wall eddy viscosity increased significantly in the LAG, which may be associated with an increased risk of postoperative anastomotic calculi and re-obstruction, whereas these parameters decreased in the MAG and SAG.

In conclusion, the MAG achieved the optimal postoperative recovery, characterized by superior DRF and post-APD improvement, significant pressure reduction, appropriate pressure gradient, mild velocity increase, and decreased wall eddy viscosity. The LAG showed marked velocity improvement and mild DRF enhancement, along with moderate pressure reduction; however, wall shear pressure and wall eddy viscosity also increased significantly, which may be related to postoperative complications. The SAG yielded the poorest postoperative outcomes.

This study represents a preliminary exploration of this novel method. Owing to the labor-intensive nature of model construction, the sample size in each group was limited. The conclusions would be more credible with a larger sample size and longer follow-up duration in future studies. In addition, due to the irregularity of human anatomical data, further validation of these results in subsequent studies and animal experiments is warranted.

Conclusion

Deep learning-based upper urinary tract model reconstruction combined with CFD simulation is a novel noninvasive method for measuring upper urinary tract urine flow dynamics parameters. Parameters including pressure, pressure gradient, velocity, and wall eddy viscosity can serve as novel indicators for evaluating postoperative recovery in pediatric hydronephrosis.

Among the three robot-assisted pyeloplasty techniques, the middle anastomosis approach yields the optimal postoperative recovery, featuring superior DRF and post-APD improvement, significant pressure reduction, appropriate pressure gradient, mild velocity increase, and decreased wall eddy viscosity. Although the large anastomosis method results in marked velocity improvement, it also leads to significant increases in wall shear pressure and wall eddy viscosity. The small anastomosis method is associated with the poorest postoperative recovery.

Abbreviations

APD

Anteroposterior Diameter

CFD

Computational fluid dynamics

CNN

Convolutional neural network

DRF

Differential renal function

GCN

Graph convolutional net work

FDA

U.S. Food and Drug Administration

LAG

Large anastomosis group

MAG

Middle anastomosis group

MRU

Magnetic resonance urography

SAG

Small anastomosis group

UPJO

Ureteropelvic junction obstruction

Author contributions

Shan Lin: Conceptualization, Methodology, Formal Analysis, Writing—Original Draft, Writing—Review & Editing, Visualization Wangbin Ding: Software, Data Curation, Validation, Investigation Jianglong Chen: Writing—Review & Editing, Validation, Investigation Yufeng He: Data Curation, Writing—Review & Editing Zhanpeng Zhu: Software, Data Curation Haoran Sun: Software, Data Curation Guangxu You: Writing—Review & Editing Yuru Zhang: Writing—Review & Editing Jingjing Lu: Writing—Review & Editing Di Xu: Project Administration, Funding Acquisition, Resources Xing Lin: Project Administration, Validation, Investigation, Supervision Shan Lin, Wangbin Ding, and Jianglong Chen contributed equally as first authors to this manuscript. Co-Corresponding Author: Di Xu (Email: fjsleyk@163.com) Corresponding Author: Xing Lin (Email: linxing@fzu.edu.cn).

Funding

This study was supported by the Leading Program of Joint Science and Technology Innovation Foundation of Fujian Province (Grant No. 2023Y9308, 2021Y9019); the Health Joint General Program of Natural Science Foundation of Fujian Province (Grant No. 2025J01072); the General Program of Qihang Foundation of Fujian Medical University (Grant No. 2022QH1300).

Data availability

The original data generated in this study are included in the article and its supplementary materials. For further inquiries, please contact the corresponding authors.

Declarations

Ethics approval and consent to participate

This study involving human participants was reviewed and approved by the Institutional Review Board of Fujian Provincial Hospital (approval no.: K2022-04–012). Written informed consent was obtained from the legal guardians of all pediatric participants. All adult participants provided written informed consent after being fully informed of the study objectives, procedures, potential risks, and benefits. This study was conducted in accordance with the ethical principles specified in the Declaration of Helsinki.

Competing interests

The authors declare no competing interests.