Magnetic resonance imaging parameter-based machine learning for prognosis prediction of high-intensity focused ultrasound ablation of uterine fibroids

Figures & data

Table 1. MRI scan parameters.

The First Affiliated Hospital of Chongqing Medical University/ Chongqing Haifu Hospital	Repetition time (ms)	Echo time (ms)	Number of excitations	Field of view (cm × cm)	Matrix size (mm × mm)	Slice thickness (mm)	Slice gap (mm)	Imaging planes
T1-weighted	175/214	1.8/10	1/1.3	40 × 28/25.2 × 36	320 × 224/320 × 70	5/5	1.5/1.0	Transverse
T2-weighted	4060/5300	100/88	3/2	22.4 × 28/24 × 24	288 × 224/320 × 75	5/5	1.5/1.0	Transverse
								sagittal
Contrast-enhanced MRI	4.2/3.94	1.9/1.84	0.72/1	38 × 30.4/35 × 28	320 × 224/288 × 75	4/2.5	0/0.5	Transverse
								sagittal
								coronal

MRI: magnetic resonance imaging.

Figure 1. MicroSea 3D Image Processing Software was used to sketch the non-perfused volume layer by layer on contrast-enhanced magnetic resonance imaging.

Figure 2. (A–E) Method for measuring the maximum thickness of residual fibroids. (F–J) Method for counting the basal distribution of residual fibroids. (F) No residual fibroid, 0; (G) limited distribution, 1 quadrant; (H) small distribution, 2 quadrants; (I) obvious distribution, 3 quadrants; (J) extensive distribution, 4 quadrants.

Table 2. Baseline characteristics of all uterine fibroids.

Baseline characteristics	Values
Total number of fibroids	573
Age (years)	39.0 (32.5–43.0)
Follow-up time (days)	203.0 (122.5–367.5)
Number of fibroids
1	260 (45.4)
2	95 (16.6)
3	35 (6.1)
4	20 (3.5)
≥5	163 (28.4)
Number of treated fibroids
1	284 (49.6)
2	95 (16.6)
3	39 (6.8)
4	23 (4.0)
≥5	132 (23.0)
FIGO type
0	16 (2.8)
1	8 (1.4)
2	4 (0.7)
3	51 (8.9)
4	261 (45.5)
5	41 (7.2)
6	157 (27.4)
7	35 (6.1)
T2WI type
I	278 (48.5)
II	72 (12.6)
III	152 (26.5)
IV	71 (12.4)
Blood supply degree
Lack of blood supply	114 (19.9)
Moderate blood supply	300 (52.4)
Rich blood supply	159 (27.7)
Maximum thickness of residual fibroids (cm)	0.8 (0.5–1.2)
Basal distribution of residual fibroids
0	56 (9.8)
1	138 (24.1)
2	170 (29.7)
3	127 (22.2)
4	82 (14.3)
NPVR	70.1% (55.8–83.1%)
Postoperative NPV (cm^3)	48.7 (21.9–90.5)
Postoperative RFV (cm^3)	19.2 (8.4–40.7)
Abdominal fat thickness (cm)	1.3 (1.0–1.8)
Maximum anteroposterior distance from the surface to the center of the fibroid	7.2 (5.8–9.0)

Data presented as n, n (%), or median (interquartile range).

FIGO: International Federation of Gynecology and Obstetrics; T2WI: T2-weighted imaging; NPV: non-perfused volume; NPVR: NPV ratio; RFV: residual fibroid volume.

Figure 3. The Boruta algorithm was used for feature screening and ranking. (A), (C) Box plots of all features attribute plus minimum, average, and maximum shadow scores; green variables were of predictive importance, yellow variables were tentative and red variables were eliminated. (B), (D) The decision process of the Boruta algorithm to accept and reject features. FIGO: International Federation of Gynecology and Obstetrics; T2WI: T2-weighted imaging; NPV: non-perfused volume; RFV: residual fibroid volume.

Figure 3. Continued

Figure 4. Comparison of the real and predicted values of NPV reduction in the testing set. (A) LiR. (B) RF. (C) MLP. LiR: linear regression; RF: random forest; MLP: multilayer perceptron; NPV: non-perfused volume.

Table 3. Performance of the three machine learning models for predicting NPV reduction.

Model		R^2	MAE	MSE	MedAE	EV
LiR	Training	0.836	10.236	347.852	6.996	0.836
	Average 10-fold CV	0.809	11.700	507.656	7.135	0.815
	Testing	0.851	13.375	526.454	7.711	0.858
RF	Training	0.982	3.121	38.537	1.334	0.982
	Average 10-fold CV	0.879	8.441	258.865	3.776	0.885
	Testing	0.828	11.380	605.767	4.957	0.828
MLP	Training	0.926	7.449	157.452	4.137	0.948
	Average 10-fold CV	0.907	7.506	191.089	4.015	0.910
	Testing	0.792	11.780	732.541	4.401	0.793

MAE: mean absolute error, MSE: mean square error, MedAE: median absolute error, EV: explained variance, LiR: linear regression, RF: random forest, MLP: multilayer perceptron, NPV: non-perfused volume, CV: cross-validation.

Figure 5. Area under the receiver operating characteristic curves of LoR, RF, and MLP to predict residual fibroid regrowth. (A) Training set results. (B) Testing set results. LoR: logistic regression; RF: random forest; MLP: multilayer perceptron.

Table 4. Performance of three machine learning models for predicting residual fibroid regrowth.

		AUC [95% CI]	Sensitivity	Specificity	Accuracy	PPV	NPV
LoR	Training	0.904 [0.877–0.923]	0.867	0.789	0.830	0.917	0.845
	Average 10-fold CV	0.883 [0.836–0.930]
	Testing	0.876 [0.831–0.917]	0.875	0.677	0.762	0.670	0.878
RF	Training	1.000 [1.000–1.000]	1.000	0.995	0.998	0.995	1.000
	Average 10-fold CV	0.904 [0.869–0.939]
	Testing	0.891 [0.850–0.929]	0.833	0.813	0.821	0.769	0.867
MLP	Training	0.847 [0.814–0.877]	0.398	0.943	0.659	0.884	0.590
	Average 10-fold CV	0.870 [0.819–0.922]
	Testing	0.771 [0.710–0.823]	0.389	0.875	0.667	0.700	0.656

LoR: logistic regression, RF: random forest, MLP: multilayer perceptron, AUC: area under the curve, PPV: positive predictive value, NPV: negative predictive value, CV: cross-validation, CI: confidence interval.

Figure 6. (A–D) A 29-year-old woman with a single uterine fibroid in the testing set. T2WI type: IV; FIGO type: 3; CE-MRI showed rich blood supply. The patient returned for MRI review 100 days after postoperative MRI evaluation. Maximum thickness of residual fibroids: 0.85 cm. Basal distribution of residual fibroids: two quadrants. Postoperative NPV: 37.3 cm³; postoperative RFV: 27.0 cm³; NPVR: 58.01%. NPV reduction value: 26.7 cm³. RFV increased by 19.63%. Predictive result of the NPV reduction value: 25.7 cm³; predictive result of residual fibroids: regrowth. (E–H) A 32-year-old woman with a single uterine fibroid in the testing set. T2WI type: III; FIGO type: 4; CE-MRI showed a lack of blood supply. The patient returned for MRI review 105 days after postoperative MRI evaluation. Maximum thickness of residual fibroids: 0.65 cm. The basal distribution of residual fibroids: two quadrants. Postoperative NPV: 332.7 cm³; postoperative RFV: 26.1 cm³; NPVR: 92.73%. NPV reduction value: 135.9 cm³. RFV decreased by 18.77%. Predictive result of NPV reduction was 103.8 cm³; predictive result of residual fibroid: non-regrowth. CE-MRI: contrast-enhanced magnetic resonance imaging; FIGO: International Federation of Gynecology and Obstetrics; T2WI: T2-weighted imaging; NPV: non-perfused volume; RFV: residual fibroid volume.

Figure 7. Prognostic prediction curves of the two cases in Figure 6. The dotted lines predict NPV values over time. The ‘•’ shows the true NPV values. The ‘+’ and ‘-’ mean regrowth and non-regrowth, respectively. (A) Residual fibroid regrowth was observed at 30, 60, 90,100, 180, 270, and 360 days after treatment, indicating continuous growth of residual fibroids; (B) The residual fibroids remained non-regrowth at 30, 60, 90, 105, 180, 270, and 360 days after treatment, showing no trend for fibroid recurrence; therefore, frequent MRI examination and excessive worry are unnecessary. NPV: nonperfused volume; MRI: magnetic resonance imaging.