Prediction of postoperative reintervention risk for uterine fibroids using clinical-imaging features and T2WI radiomics before high-intensity focused ultrasound ablation

Abstract

Objective

To predict the risk of postoperative reintervention for uterine fibroids using clinical-imaging features and T2WI radiomics before high-intensity focused ultrasound (HIFU) ablation.

Methods

Among patients with uterine fibroids treated with HIFU from 2019 to 2021, 180 were selected per the inclusion and exclusion criteria (42 reintervention and 138 non-reintervention). All patients were randomly assigned to either the training (n = 125) or validation (n = 55) cohorts. Multivariate analysis was used to determine independent clinical-imaging features of reintervention risk. The Relief and LASSO algorithm were used to select optimal radiomics features. Random forest was used to construct the clinical-imaging model based on independent clinical-imaging features, the radiomics model based on optimal radiomics features, and the combined model incorporating the above features. An independent test cohort of 45 patients with uterine fibroids tested these models. The integrated discrimination index (IDI) was used to compare the discrimination performance of these models.

Results

Age (p < .001), fibroid volume (p = .001) and fibroid enhancement degree (p = .001) were identified as independent clinical-imaging features. The combined model had AUCs of 0.821 (95% CI: 0.712–0.931) and 0.818 (95% CI: 0.694–0.943) in the validation and independent test cohorts, respectively. The predictive performance of the combined model was 27.8% (independent test cohort, p < .001) and 29.5% (independent test cohort, p = .001) better than the clinical-imaging and radiomics models, respectively.

Conclusion

The combined model can effectively predict the risk of postoperative reintervention for uterine fibroids before HIFU ablation. It is expected to help clinicians to develop accurate, personalized treatment and management plans. Future studies will need to be prospectively validated.

Keywords:

• Uterine fibroid
• high-intensity focused ultrasound
• radiomics
• reintervention
• prediction

Introduction

Uterine fibroids are common benign gynecological tumors [1]. According to a study of randomly selected American women aged 35–49, the cumulative incidence of fibroids among white women was nearly 70% [2]. Uterine fibroids can cause severe clinical symptoms, including abnormal bleeding, excessive menstruation, anemia and dysmenorrhea which require prompt treatment [3]. Although there are several uterine-sparing treatment options, such as myomectomy and uterine artery embolization (UAE), high-intensity focused ultrasound (HIFU) has unique advantages in increasing pregnancy rate and speed of postoperative recovery [4,5]. Therefore, it has been widely used to treat uterine fibroids, primarily relieving or eliminating the clinical symptoms associated with fibroids [6,7]. HIFU treatment is safe, effective and noninvasive. However, HIFU is associated with a higher risk of reintervention than other uterine-sparing treatment options [8,9]. Reintervention rates vary with follow-up time. Xu et al. reported that the 24-month reintervention rates of myomectomy, UAE and HIFU in treating uterine fibroids were 10%, 8% and 14%, respectively. The 36-month reintervention rates were 9%, 14% and 22%, respectively. The reintervention rate of HIFU increased rapidly after 60 months [10]. Reintervention imposes an additional physical and financial burden on patients. Thus, predicting the risk of postoperative reintervention before HIFU ablation would help clinicians develop individualized treatment plans to reduce reintervention rates or determine early management plans for chronic disease, which will sustain relief of postoperative clinical symptoms and improve patients’ long-term quality of life. Unfortunately, there are no reliable quantitative tools for predicting the risk of postoperative reintervention.

Magnetic resonance imaging (MRI) is a conventional examination method for detecting uterine fibroids. Radiomics is a powerful technique that rapidly extracts many quantitative features from tomographic images through high-throughput computing, transforming digital medical images into mineable, high-dimensional data for improved decision support [11,12]. Radiomics has been used to differentiate uterine sarcomas from atypical fibroids and to predict the non-perfusion volume rate (NPVR) after HIFU ablation [13,14]. However, no radiomics study has explored the prediction of reintervention risk after HIFU so far. Therefore, this study aimed to predict the risk of postoperative reintervention for uterine fibroids using clinical-imaging features and T2-weighted imaging (T2WI) radiomics before HIFU ablation. In so doing, we hope to provide clinicians with a reliable imaging tool for developing accurate, personalized treatment and management plans.

Materials and methods

Patients

The Institutional Review Board approved this retrospective cohort study, and the requirement for informed consent was waived (IRB No. 2023ER214-1). We consecutively collected 265 patients with uterine fibroid treated by HIFU at our hospital between January 2019 and January 2021.

The inclusion criteria were as follows [15–17]: (1) premenopausal women over 18 years of age with a clinical and imaging diagnosis of uterine fibroids; (2) first-time HIFU ablation treatment; (3) HIFU-treated single fibroid or the largest in multiple fibroids to facilitate follow-up surveys; (4) MRI examination performed within 3 days before and after HIFU ablation; (5) no previous history of surgery or medication. The exclusion criteria were as follows [17–19]: (1) the MRI sequences were incomplete, or some images were missing, which might result in an inability to segment the tumor; (2) MR image quality was poor, which might affect the accuracy of feature extraction; (3) missed visits for various reasons, which might result in an inability to trace reintervention information; (4) reintervention of non-fibroid factors, such as the patient’s psychological factors, fertility requirements, other surgical indications, suspected uterine sarcoma, etc., which might affect the accuracy of grouping; (5) symptomatic recurrence and/or increased fibroid volume, but no decision has been made to undergo reintervention. The decision to undergo reintervention was also related to the local level of medical care, the patient’s economic condition and educational attainment. These heterogeneities might affect the accuracy of the study results. Finally, 180 eligible patients (median age 43.5; age range 23–64) were included in the study (Figure 1(A)), with a median follow-up time of 33 months (interquartile range: 25–41). All patients were randomly divided at a 7:3 ratio to the training cohort (125 patients, 29 reintervention; 96 non-reintervention) and the validation cohort (55 patients, 13 reintervention; 42 non-reintervention). Using the same inclusion and exclusion criteria, 45 patients with uterine fibroids (median age 43; age range 25–62) treated by HIFU at our hospital between June 2017 and December 2018 were retrospectively collected to comprise an independent test cohort (12 reinterventions; 33 non-reinterventions; Figure 1(B)). The median follow-up was 52 months (interquartile range: 21–60).

Figure 1. Flowchart of patient recruitment (A and B).

MR image acquisition

MRI examinations were performed using a 3.0 T MR scanner (Discovery 750; GE Healthcare) with a 16-channel phased-array body coil. Before the examination, patients were trained to breathe shallowly and slowly. Their abdomen was compressed using a bandage to reduce the effect of respiratory motion artifacts on the accuracy of tumor segmentation and feature extraction. All patients underwent MRI within 3 days before and after ablation. The scan sequences included sagittal fat-suppressed T2-weighted imaging (T2WI-FS) and sagittal contrast-enhanced T1-weighted imaging (CE-T1WI) sequences. The specific scan parameters are provided in Supplementary Table S1. For the CE-T1WI sequence, GD-DTPA (Magnevist, Bayer Schering Pharma, Berlin, Germany) was intravenously injected at a 0.1 mmol/kg dose and a flow rate of 1.0 ml/s. Late arterial images were acquired approximately 45 s after GD-DTPA injection to analyze the clinical-imaging features.

Follow-up survey

Patients were followed up by telephone to obtain information about the interval between reintervention, the number of reinterventions, the reason for reintervention and the method of reintervention. Reintervention was defined as any additional intervention for fibroid treatment after HIFU ablation due to unrelieved symptoms, recurrence of symptoms, or increased fibroid volume [20,21]. Non-reintervention was defined as no treatment with any additional intervention for fibroid after HIFU ablation until the follow-up cutoff date (2023.2). Reasons for reintervention included unrelieved symptoms, recurrence of symptoms, increased fibroid volume, fertility requirements, other indications for gynecologic surgery, psychological factors and suspected uterine sarcoma. Reintervention options included laparoscopic/hysteroscopic myomectomy, UAE, hysterectomy, HIFU and pharmacological treatment. Follow-up duration was calculated as the time from HIFU treatment to the time of reintervention or study cutoff (2023.2).

Clinical-imaging features

Two radiologists, Y.Z. and S.Q., with 5 years and 4 years of experience in gynecologic MRI, independently evaluated the clinical-imaging features using a double-blind method. In case of disagreement, they reached a consensus by discussion. The clinical features included age, BMI and peripheral platelet count. The imaging features had uterine and fibroid location, fibroid volume, fibroid type, fibroid central layer abdominal wall thickness, distance from the ventral surface of the fibroid to the skin, fibroid T2WI-FS signal intensity (hypointense: equivalent to skeletal muscle; isointense: higher than skeletal muscle but lower than myometrium; hyperintense: similar to or higher than myometrium) [22], fibroid enhancement degree (mild: lower than myometrium; moderate: similar to myometrium; obvious: higher than myometrium) [22], homogeneity of the fibroid enhancement signal, and fibroid number. The HIFU treatment time, projection energy and NPVR (non-perfusion volume/fibroid volume) were recorded.

Tumor segmentation and feature extraction

A radiologist (Q.J. with 5 years of experience in MRI, reader 1) used open-source software (3D Slicer, Version 4.10.2, https://www.slicer.org) to manually segment the region of interest (ROI) along the edge of the fibroid layer by layer on T2WI-FS images. The ROI of each layer was fused to generate a three-dimensional region of interest volume (VOI) of the fibroid. The feature extraction process in our study followed the Image Biomarker Standardization Initiative (IBSI) guidelines. The features of each VOI were extracted from the original image and from the transformed image after Laplacian Gaussian filtering and wavelet transform filtering. A total of 1223 radiomics features were extracted, including first-order features, shape features, gray-level size zone matrix (GLSZM), gray-level dependence matrix (GLDM), neighboring gray-tone difference matrix (NGTDM), gray-level run-length matrix (GLRLM), gray-level co-occurrence matrix (GLCM) and filter-derived features.

Another radiologist (S.Q. with 4 years of experience in gynecologic MRI, reader 2) randomly selected 1/3 of the images for ROI segmentation and extracted features. The reproducibility of the radiomics features extracted by the two radiologists (reader 1, reader 2) was evaluated using the intergroup correlation coefficient (ICC). Features with an ICC >0.75 were considered reliable and proceeded to the following analysis step.

Data pre-processing and feature selection

The smote algorithm was used to oversample the minority (reintervention) class and undersample the majority (non-reintervention) class in the training cohort to achieve class balance. The smote algorithm was implemented using the ‘DMwR’ package in the R software (version 4.1.1). After the sampling, the minority (reintervention) class was five times larger than before; the majority (non-reintervention) class was four times larger than the pre-sampling minority (reintervention) class. This approach has resulted in better classifier performance than undersampling the majority class alone [23].

R software (version 4.1.1, https://www.r-project.org) and the uAI Research Portal software (version 730) were used to perform feature selection. Before feature selection, Z-score normalization was applied to remove the effect of dimensionality between features. The Relief algorithm and the least absolute shrinkage and selection operator (LASSO) algorithm were used to identify and select the optimal radiomics features in the training cohort. First, the Relief algorithm was used to assign different weights to the features based on their relevance to each class. Features with weights less than a threshold value were then removed. Next, the LASSO algorithm was used to select the most valuable subset of features. In the LASSO algorithm, the optimal value of the penalty factor λ was chosen according to 10-fold cross-validation.

Model construction

The clinical-imaging model was constructed using independent clinical-imaging predictors. For the radiomics and combined models, linear discriminant analysis (LDA) was used to find the best projection line for the optimal prediction features. The aim was to maximize the distance between data points of different classes (reintervention/non-reintervention) and minimize the distance between data points of the same class. LDA has some classification ability, and its results were used as input features for the random forest to construct models to more accurately identify patients with and without reintervention, with a classification threshold of 0.5.

Model evaluation and validation

Receiver operating characteristic curves (ROC) were plotted to analyze the predictive performance of the models, and the areas under the ROC curves (AUC), sensitivity, specificity, accuracy and F1-score were calculated (The formulae for calculating performance metrics are shown in Supplementary Table S3). The net reclassification index (NRI) was used to assess the predictive accuracy improvement of the new model compared to the old model. The integrated discrimination index (IDI) was utilized to evaluate the overall improvement in the predictive performance of the new model over the old model. The DeLong test was utilized to compare the predictive performance between different models. The Brier score was employed to evaluate the error between the model’s predicted probabilities and the actual values. Decision curve analysis (DCA) was conducted to assess whether the models contribute to developing a clinical treatment strategy by quantifying the net benefit of the models at different risk threshold probabilities. Figure 2 illustrates the workflow of the radiomics analysis.

Figure 2. Flowchart of radiomics analysis.

Statistical analysis

Kolmogorov–Smirnov (K-S) test was used for the normality test for continuous variables with a sample size >50, and Shapiro–Wilk (S-W) test was utilized for the normality test for continuous variables with a sample size <50. Continuous variables conforming to the normal distribution were presented as mean ± standard deviation (x̅ ± s), while those that did not conform were described using the quartile method [M (Q1, Q3)]. For normally distributed continuous variables where the two groups were independent, the independent samples t-test was used to compare the differences between the two groups of continuous variables. For non-normal distributed continuous variables where the two groups were independent of each other, the Mann–Whitney U test was employed to compare differences between the two groups of continuous variables. For categorical variables, the chi-square test or a continuity-corrected chi-square test was utilized to compare differences between the two groups. Multivariate analysis was conducted to identify independent clinical-imaging features that predicted the risk of reintervention. Statistical analyses were performed using SPSS Statistics version 26.0 and R software version 4.1.1. Two-sided p values less than .05 were considered statistically significant.

Results

Reintervention follow-up results

In the training and validation cohorts, 226 patients were included in the reintervention follow-up survey (excluding those with incomplete MRI sequences or missing images and poor MR image quality) at a median follow-up time of 33 months (interquartile interval: 25–41). Of these, 36 patients missed follow-ups for various reasons, and 6 were re-intervened for non-fibroid factors (1 had fertility requirements, 2 had other surgical indications, and 3 had psychological factors). Four patients presented with symptomatic recurrence and/or increased fibroid volume, but no decision has been made to re-intervene (Figure 1(A)). Among the remaining 180 patients, 23.3% (42/180) underwent reintervention treatment. The median reintervention interval post-HIFU was 24 months (interquartile range: 13–33). Of these, 35 (83.3%) received hysteroscopic/laparoscopic myomectomy, 4 (9.5%) received a hysterectomy, and 3 (7.2%) underwent pharmacological treatment. Reasons for reintervention included symptom recurrence in 6 (14.3%) patients, increased fibroid volume in 18 (42.8%) patients, increased fibroid volume combined with symptom recurrence in 5 (11.9%) patients, and unrelieved symptoms in 13 (31.0%) patients. In the independent test cohort, 61 patients were included in the reintervention follow-up survey with a median follow-up time of 52 months (interquartile interval: 21–60). Among them, 9 patients missed follow-ups for various reasons, and 4 were re-intervened due to non-fibroid factors (2 had other surgical indications, 2 had psychological factors). Three patients presented with symptomatic recurrence and/or increased fibroid volume, but no decision has been made to re-intervene (Figure 1(B)). Among the remaining 45 patients, 26.7% (12/45) underwent reintervention treatment. The median reintervention interval post-HIFU was 21 months (interquartile range: 7–38). Of these, 6 (50%) received hysteroscopic/laparoscopic myomectomy, and 6 (50%) received pharmacological treatment. Reasons for reintervention included symptom recurrence in 1 (8.3%) patient, increased fibroid volume in 6 (50%) patients, increased fibroid volume combined with symptom recurrence in 2 (16.7%) patients, and unrelieved symptoms in 3 (25.0%) patients.

Clinical-imaging features

Finally, there were 180 people in the training cohort and validation cohort of this study. The differences in clinical-imaging features between the reintervention group and the non-reintervention group are shown in Table 1. Age, fibroid volume, T2WI-FS signal intensity and fibroid enhancement degree showed significant differences. HIFU treatment results differed significantly regarding treatment time, projection energy and NPVR. Patients in the reintervention group were younger than those in the non-reintervention group (p < .001). The fibroid volume of the reintervention group was larger than that of the non-reintervention group (p = .032). The T2WI-FS signal intensity was predominantly hyperintense in the reintervention group and hypointense in the non-reintervention group (p = .002). The percentage of obvious enhancement in the reintervention group was noticeably higher than in the non-reintervention group. (p = .002). The reintervention group had a longer treatment time (p < .001) and higher projection energy (p < .001) than the non-reintervention group; however, the reintervention group had a lower NPVR (p < .001) than the non-reintervention group. Multivariate analysis revealed that age, fibroid volume and enhancement degree were independent predictors of reintervention risk post-HIFU (Table 2, Figure 3). The differences in clinical-imaging features between the reintervention and non-reintervention groups for the independent test cohort are shown in Table 3. T2WI-FS signal intensity, fibroid enhancement degree, and NPVR showed significant differences.

Figure 3. A 40-year-old patient with a uterine fibroid (white arrow). The fibroid volume before ablation was 181,567.960 mm³, and the residual fibroid volume after ablation was 76,321.690 mm³. The NPVR value was 58.0%. Thirty months after HIFU ablation, this patient underwent a repeat hysteroscopic myomectomy for lack of symptom relief. (A) Before HIFU ablation, the fibroid showed homogeneous hyperintense on the T2WI-FS image. (B) The fibroid showed obvious homogeneous enhancement on the late arterial image. (C) Late arterial images after HIFU ablation showed no enhancement in the ablated area.

Table 1. Clinical-imaging features of the training and validation cohorts.

	Reintervention grouping
	Non-reintervention	Reintervention
Clinical-imaging features	(n = 138)	(n = 42)	p
Age (years)^a	44 (40–48)	41 (35–44)	<.001^b
BMI (kg/m^2)^a	22.170 (21.025–23.351)	22.222 (20.860–24.061)	.655^b
Peripheral platelet count (10 e^9/L)^a	223.500 (180.750–286.500)	236.000 (185.000–294.500)	.572^b
Uterine fibroid location			.220^c
Anterior	90 (65.2)	23 (54.8)
Posterior	48 (34.8)	19 (45.2)
Uterine position			.633^c
Anterior	100 (72.5)	32 (76.2)
Posterior	38 (27.5)	10 (23.8)
Volume (mm^3)^a	60,738.160 (38,579.282–105,838.971)	89,745.900 (43,435.293–140,692.974)	.032^b
Type			.897^c
Subserous	9 (6.5)	2 (4.8)
Intramural	118 (85.5)	37 (88.1)
Submucous	11 (8.0)	3 (7.1)
The thickness of the abdominal wall (mm)^a	28.465 (21.700–33.905)	26.404 (20.650–32.553)	.262^c
Distance from the ventral surface of the fibroid to the skin (mm)^a	50.750 (34.045–70.558)	50.600 (31.130–68.365)	.653^c
T2WI signal intensity			.002^c
Hypointense	79 (57.2)	11 (26.2)
Isointense	24 (17.4)	11 (26.2)
Hyperintense	35 (25.4)	20 (47.6)
Enhancement degree
Mild	61 (44.2)	13 (31.0)	.002^c
Moderate	58 (42.0)	13 (31.0)
Obvious	19 (13.8)	16 (38.0)
Enhancement signal homogeneity
Homogeneous	31 (22.5)	13 (31.0)	.262^c
Inhomogeneous	107 (77.5)	29 (69.0)
Fibroid number^a	1 (1–1)	1 (1–1)	.488^b
Treatment time (s)^a	485.500 (327.000–694.000)	725.500 (532.000–1234.750)	<.001^b
Projection energy (J)^a	190,800 (123,500–277,900)	289,300 (213,700–493,900)	<.001^b
NPVR (%)^a	80.6 (66.2–93.8)	58.6 (34.5–74.3)	<.001^b

^aData are medians and quartiles in parentheses.

^bp Values were obtained using the Wilcoxon rank-sum test.

^cp Values were obtained using the chi-square test.

Unless otherwise stated, data are shown as the number of patients and percentage in parentheses.

Table 2. Independent predictors for reintervention risk.

Parameters	Regression coefficient	p	Odds ratio (95%CI)
Age	−0.129	<.001	0.879 (0.945–0.940)
Volume	0.000	.001	1.000 (1.000–1.000)
Mild enhancement	−1.704	.001	0.182 (0.067–0.492)
Moderate enhancement	−1.595	.002	0.203 (0.074–0.554)

Table 3. Clinical-imaging features of the independent test cohort.

	Reintervention grouping
	Non-reintervention	Reintervention
Clinical-imaging features	(n = 33)	(n = 12)	p
Age (years)^a	43 (40–46)	41 (31–46)	.189^b
BMI (kg/m^2)	21.172 ± 2.182	21.639 ± 1.438	.496^c
Peripheral platelet count (10 e^9/L)	218.940 ± 76.477	234.250 ± 65.247	.541^c
Uterine fibroid location			1.000^d
Anterior	23 (69.7)	8 (66.7)
Posterior	10 (30.3)	4 (33.3)
Uterine position			.344^d
Anterior	24 (72.7)	11 (91.7)
Posterior	9 (27.9)	1 (8.3)
Volume (mm^3)^a	45,742.349 (26,983.192–80,345.535)	54,637.077 (40,799.978–90,402.222)	.317^b
Type			.685^d
Subserous	0 (0.0)	0 (0.0)
Intramural	30 (90.9)	12 (100)
Submucous	3 (9.1)	0 (0.0)
The thickness of the abdominal wall (mm)	24.936 ± 6.518	28.523 ± 9.849	.164^c
Distance from the ventral surface of the fibroid to the skin (mm)^a	47.460 (32.990–91.500)	55.840 (32.253–82.813)	.878^b
T2WI signal intensity			<.001^d
Hypointense	21 (63.6)	3 (25.0)
Isointense	9 (27.3)	1 (8.3)
Hyperintense	3 (9.1)	8 (66.7)
Enhancement degree
Mild	24 (72.7)	5 (41.7)	.041^d
Moderate	8 (24.2)	4 (33.3)
Obvious	1 (3.1)	3 (25.0)
Enhancement signal homogeneity			1.000^d
Homogeneous	5 (15.1)	2 (16.7)
Inhomogeneous	28 (84.9)	10 (83.3)
Fibroid number^a	1 (1–2)	1 (1–2)	.414^b
Treatment time (s)	534.667 ± 216.543	882.167 ± 664.218	.101^c
Projection energy (J)	210,366.667 ± 87,757.210	371,666.667 ± 263,255.511	.060^c
NPVR (%)	88.0 ± 25.5	65.9 ± 39.5	.033^c

^aData are medians and quartiles in parentheses.

^bp Values were obtained using the Wilcoxon rank-sum test.

^cp Values were obtained using the independent samples t-test.

^dp Values were obtained using the chi-square tests for continuity correction.

Unless otherwise stated, data are shown as the number of patients and percentage in parentheses.

Radiomics feature selection and model evaluation

Among the 1223 radiomics features extracted, 798 demonstrated good inter-observer agreement (ICC > 0.75). The Relief and LASSO algorithm were utilized to select the optimal radiomics features, resulting in nine retained features used to construct the models (Supplementary Table S2). The AUC of the clinical-imaging model based on age, fibroid volume and fibroid enhancement degree was 0.757 (95% confidence interval [CI]: 0.700–0.815) in the training cohort and 0.722 (95% CI: 0.591–0.852) in the validation cohort. The AUC of the radiomics model based on the optimal radiomics features was 0.785 (95% CI: 0.731–0.840) in the training cohort and 0.734 (95% CI: 0.571–0.898) in the validation cohort. The combined model incorporating clinical-imaging and radiomics features showed excellent and consistent predictive performance, with AUC of 0.842 (95% CI: 0.796–0.888) and 0.821 (95% CI: 0.712–0.931) for the training and validation cohorts, respectively. The combined model also showed the best predictive performance in the independent test cohort, with an AUC of 0.818 (95% CI: 0.694–0.943; Table 4, Figures 4 and 5). The IDI results showed an overall improvement of 20.8% (validation cohort, p < .001) and 27.8% (independent test cohort, p < .001) in predictive performance for the combined model compared to the clinical-imaging model. Additionally, the combined model demonstrated an overall improvement of 22.0% (validation cohort, p = .004) and 29.5% (independent test cohort, p = .001) compared to the radiomics model (Table 5). The DeLong test revealed that the combined model outperformed the clinical-imaging model regarding predictive performance (training cohort, p = .004; validation cohort, p = .258; independent test cohort, p = .07).

Figure 4. Receiver operating characteristic (ROC) curves of the clinical-imaging mode, radiomics model and combined model in the training (A) and validation (B) cohorts.

Figure 5. Receiver operating characteristic (ROC) and precision-recall curves of the combined model in the independent cohort (A and B).

Table 4. Predictive performance of the clinical-imaging, radiomics and combined models.

		Training			Validation			Test
Model	Clinical-imaging	Radiomics	Combined	Clinical-imaging	Radiomics	Combined	Clinical-imaging	Radiomics	Combined
AUC (95% CI)	0.757 (0.700–0.815)	0.785 (0.731–0.840)	0.842 (0.731–0.888)	0.722 (0.591–0.852)	0.734 (0.571–0.898)	0.821 (0.712–0.931)	0.670 (0.495–0.846)	0.668 (0.450–0.886)	0.818 (0.694–0.943)
Sensitivity	0.648	0.774	0.800	0.692	0.444	0.923	0.333	0.667	0.667
Specificity	0.707	0.819	0.672	0.690	0.881	0.619	0.786	0.805	0.825
Accuracy	0.674	0.640	0.743	0.691	0.745	0.691	0.622	0.733	0.756
F1-score	0.689	0.605	0.776	0.541	0.364	0.585	0.320	0.571	0.593

Table 5. Comparison of the model’s predictive performance.

Model	Cohort	NRI (95% CI)	p (NRI)	IDI (95% CI)	p (IDI)
Combined model vs. Clinical-Imaging model	Training	0.117 (0.001–0.2335)	.048	0.154 (0109–0.199)	<.001
	Validation	0.159 (−0.116–0.435)	.257	0.208 (0.123–0.293)	<.001
	Independent test	0.394 (0.079–0.709)	.014	0.278 (0.148–0.408)	<.001
Combined model vs. Radiomics model	Training	0.157 (0.014–0.300)	.031	0.133 (0.068–0.197)	<.001
	Validation	0.354 (0.036–0.670)	.029	0.220 (0.07–0.370)	.004
	Independent test	0.258 (−0.111–0.626)	.170	0.295 (0.114–0.477)	.001

NRI: net reclassification index; IDI: integrated discrimination improvement.

If NRI/IDI > 0, it is an improvement, indicating that the new model has improved its predictive ability compared to the old model; if NRI/IDI < 0, it is a negative improvement, indicating that the predictive performance of the new model has decreased; if NRI/IDI = 0, it means that the predictive performance of the new model has not changed.

Clinical application

The combined model demonstrated better predictive accuracy with a Brier score of 0.174 in the validation cohort (Figure 6(A,B)). The decision curves indicated that the combined model yielded a more significant net benefit over a broader range of threshold probabilities than the clinical-imaging model, radiomics model, all-treatment and no-treatment (Figure 6(C)).

Figure 6. Evaluation and verification of the combined model. (A, B) calibration curves for the combined model regarding the agreement between predicted and actual reintervention risk in training (A) and validation (B) cohort. The x-axis represents the probability predicted by the combined model, and the y-axis represents the actual probability. The perfect prediction corresponds to the thick black dashed line. The thin black dashed line indicates the entire cohort, and the solid black line is bias-corrected by bootstrapping (1000 repetitions), showing the observed combined model performance. (C) Decision curve analysis for the combined model. The thick black line represents the net benefit of assuming no fibroid patients receive HIFU ablation. The thin black line is the net benefit of taking that all fibroid patients receive HIFU ablation. The green, blue and red lines represent the expected net benefit of predicting the risk of reintervention using the clinical-imaging, radiomics and combined models.

Discussion

The reintervention rate in this study was found to be 13.8% (31/225) at 24 months and 23.1% (52/225) at 48 months following HIFU. The primary cause for reintervention was the observed increase in fibroid volume, consistent with the findings of Zhou et al. [24]. Previous studies have emphasized the second to fourth-year post-HIFU as a crucial follow-up period [19]. Therefore, we focused on patients treated with HIFU between January 2019 and January 2021. However, longer follow-up is necessary to improve the model’s applicability. Thus, this study retrospectively collected patients treated with HIFU between June 2017 and December 2018 to comprise an independent test cohort to test the combined model’s generalization ability. The combined model, which incorporated clinical-imaging and T2WI radiomics features, achieved the best and most stable predictive performance. The overall predictive performance of the combined model was significantly better than that of the clinical-imaging and radiomics models. This model can assist clinicians in identifying high-risk groups for reintervention before HIFU ablation to develop more accurate and appropriate treatment strategies or to take measures early after ablation, such as regular exercise [25], vitamin supplementation [26] and endocrine regulation to achieve a sustained reduction in fibroid volume and clinical symptoms.

This study found that younger women were more likely to undergo reintervention post-HIFU, consistent with the results of Li et al. [9]. The growth of fibroids depends on estrogen and progesterone, and high levels of these hormones in young women may increase residual fibroid volume, leading to the recurrence of symptoms [27]. Conversely, as women reach menopause, fibroids begin to shrink since the secretion of estrogen and progesterone decreases [3]. Young women may therefore be at higher risk for reintervention post-HIFU. This study also found that fibroid volume was significantly larger in the reintervention group than in the non-reintervention group. This may be due to the larger fibroids having more tissue and blood supply vessels [28]. Additionally, the percentage of hyperintense and obvious enhancing fibroids was higher in the reintervention group than in the non-reintervention group, which is consistent with the findings of Liu et al. [28]. The hyperintense fibroids contain abundant and closely arranged smooth muscle cells, with few fibers in the extracellular matrix [29,30]. These factors reduce ultrasound energy deposition in the target area, decreasing efficacy. The abundant blood flow within fibroids takes away the thermal energy of ultrasound, resulting in a temperature within the target area lower than the ablation set temperature, leading to a decrease in NPVR [31]. The more residual tissue that remains after ablation, the higher the likelihood of regrowth and, therefore, the greater the chance of reintervention [24]. However, evaluating clinical-imaging features primarily relies on physicians’ image-reading qualifications and subjective judgment, introducing uncertainties in the prediction results of the clinical-imaging model.

Numerous studies have demonstrated that quantitative radiomics features of MR images can effectively describe the difficulty of HIFU ablation for uterine fibroids [16,22]. This suggests that radiomics can reflect the tissue heterogeneity within different NPVR fibroids. T2WI is a crucial sequence in the conventional MR scanning protocol for uterine fibroids, providing good contrast resolution without additional contrast agents and reducing the burden on the patient. Therefore, this study attempted to use T2WI radiomics before HIFU ablation to identify patients at high risk for reintervention after HIFU ablation. However, as the post-HIFU reintervention rate is typically around 20%, this may result in data class imbalance. If the difference in sample size between the classes in the training cohort is too significant, it can cause substantial deviation in the prediction accuracy of different classes, leading to a bias toward the class with more samples [23,32]. To address this problem, the study oversampled the reintervention class in the training cohort using the smote algorithm, ensuring a ratio of reintervention to non-reintervention classes close to 1:1 to improve the model’s classification accuracy. Qian et al. [33] reported that LDA could potentially achieve the best discrimination between high and low Ki67 status of intrahepatic cholangiocarcinoma. LDA is a linear, supervised feature dimensionality reduction method that takes prior knowledge about the classes under investigation [34]. On the other hand, unsupervised feature dimensionality reduction methods, like principal component analysis (PCA), cannot use classes’ prior knowledge. After performing LDA dimensionality reduction, the difference between the discrimination of reintervention and non-reintervention was maximized, resulting in more accurate patient discrimination. Six wavelet features were included in the radiomics features extracted in this study. Wavelet features can capture tumor tissue’s spatial heterogeneity at multi-scales (multi-resolution), providing more detailed information on tumor biology [35]. Among these, wavelet-HHL-GLCM-IMC1 was selected as the crucial feature using the LASSO algorithm. Larger IMC1 values indirectly indicate higher tissue homogeneity [36]. In the present data, the IMC1 values in the reintervention group were larger than those in the non-reintervention group, which might be associated with the more homogeneous fibroid tissue in the reintervention group. Zhao et al. [37] used hematoxylin–eosin and Sirius scarlet staining to observe tightly arranged smooth muscle cells and less degenerative tissue in homogeneous hyperintense fibroids. Additionally, immunohistochemical results demonstrated high expression of estrogen and progesterone receptors in homogeneous hyperintense fibroids, essential factors associated with post-HIFU reintervention. Although the predictive performance of the radiomics model was not statistically different from that of the clinical-imaging model, it was based on quantitative histological features. Radiomics features can provide detailed information on the structural nuances of the tissue, such as extracellular matrix composition, cell proliferation and neovascularisation [38,39], which are associated with fibroid reintervention mechanisms.

Previous studies have suggested a complementary value between clinical-imaging and radiomics features [22]. Hence, this study constructed a combined model incorporating clinical-imaging features and radiomics features. The results revealed that the combined model achieved excellent predictive performance in the independent test cohort, with an AUC of 0.818. This result indicated that the combined model had some generalization ability and could be applied in different clinical settings. And the combined model significantly improved predictive accuracy by 27.8% and 29.5% in the independent test cohort compared to the clinical-imaging and radiomics models, respectively. The decision curves also revealed that the combined model could achieve the highest net benefit at most range risk thresholds, suggesting that the combined model had specific clinical value. Zhou et al. [24] also demonstrated that the combined model exhibited excellent predictive performance in predicting residual fibroid regeneration after HIFU ablation.

This study has some limitations. First, it was a retrospective investigation in which only single fibroids or the largest in multiple fibroids were included. However, existing studies have not yet identified specific differences in the pathogenesis, prognosis, outcome, tissue components, or pathophysiological basis between single and multiple fibroids. Consequently, the combined model also holds potential utility in patients with multiple fibroids. Second, some patients with incomplete information were excluded, which may have introduced selection bias. While selection bias is inevitable in studies with relatively small sample sizes, this study employed a continuous patient collection approach for individuals with uterine fibroids. A small number of patients who did not meet the study’s eligibility criteria were subsequently excluded, which might have had less impact on the accuracy of the results. In future, we will further expand the sample sizes to minimize the effect of selection bias on the results. Third, only internal validation was performed because the number of patients requiring reintervention was limited. However, to further test the generalization capability of the model, an independent test cohort was established in this study. Additionally, future research endeavors will focus on external validation of the model’s predictive performance. Finally, while the literature reports that reintervention post-HIFU occurs between years 2 and 4, longer follow-up is necessary to broaden the model’s applicability. With the widespread use of HIFU in clinical practice, we will continue to follow up on patients with fibroids treated by HIFU to analyze their reintervention information.

In conclusion, this study indicated that the combined model incorporating clinical-imaging features and T2WI radiomics could effectively predict the risk of postoperative reintervention for uterine fibroids before HIFU ablation. The model has the potential to assist clinicians in formulating accurate and individualized treatment plans to reduce reintervention rates. Furthermore, it may assist in developing early chronic disease management plans following HIFU ablation, leading to a sustained reduction in fibroid volume and clinical symptoms and ultimately improving patients’ long-term quality of life.

Acknowledgements

The authors thank all the reviewers who participated in the review, as well as MJEditor (www.mjeditor.com) for providing English editing services during the preparation of this manuscript.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data used to support the findings of this study are available from the corresponding author upon request.

Additional information

Funding

This study has received funding from Nanchong City School Cooperation Research Project [No. 19SXHZ0429].