cs.CVMar 6, 2026
SaveArtificial Intelligence for Detecting Fetal Orofacial Clefts and Advancing Medical Education
Organizations: National-Regional Key Technology Engineering Laboratory for Medical Ultrasound, School of Biomedical Engineering, Shenzhen University Medical School, Shenzhen University, Shenzhen 518060, China. · Medical Ultrasound Image Computing (MUSIC) Laboratory, Shenzhen University, Shenzhen 518060, China. · Marshall Laboratory of Biomedical Engineering, Shenzhen University, Shenzhen 518060, China. · Shenzhen Luohu People’s Hospital, Shenzhen 518001, China. ·
Abstract
Orofacial clefts are among the most common congenital craniofacial abnormalities, yet accurate prenatal detection remains challenging due to the scarcity of experienced specialists and the relative rarity of the condition. Early and reliable diagnosis is essential to enable timely clinical intervention and reduce associated morbidity. Here we show that an artificial intelligence system, trained on over 45,139 ultrasound images from 9,215 fetuses across 22 hospitals, can diagnose fetal orofacial clefts with sensitivity and specificity exceeding 93% and 95% respectively, matching the performance of senior radiologists and substantially outperforming junior radiologists. When used as a medical copilot, the system raises junior radiologists' sensitivity by more than 6%. Beyond direct diagnostic assistance, the system also accelerates the development of clinical expertise. A pilot study involving 24 radiologists and trainees demonstrated that the model can improve the expertise development for rare conditions. This dual-purpose approach offers a scalable solution for improving both diagnostic accuracy and specialist training in settings where experienced radiologists are scarce.
Figures & tables
Figure 1: Design of the AIOC system. (a) Development of the AIOC system. It is based on a dual-branch network, integrates detection (YOLOX) and classification (MILA) branches for accurate diagnosis of OC and healthy cases (i.e., CL, CLP, and Control). It also visualizes view classification (i.e., NLV, CLV, NAPV, CAPV) and key structures (i.e., upper lip (purple), alveolar ridge (yellow), cleft lip (blue), cleft alveolus (red), and cleft palate (green)). (b) The internal dataset, OC-6000. (c) The external datasets, OC-GT3000 and OC-Early. (d) Case-based quantitative evaluation of AIOC and a comparative experiment with radiologists. CL: cleft lip; CLP: cleft lip and palate; CLV: cleft lip view; NLV: normal lip view; NAPV: normal alveolus and palate view; CAPV: cleft alveolus and palate view.
Figure 2: Medical training pilot study design. (a) Overview of the pilot study workflow with theoretical lecture, randomized allocation to AI-augmented or traditional training groups, and four training-and-exam cycles. (b) AI-augmented training using 20 cases from OC-6000 with AIOC system assistance. The participants will be provided with additional view categories, bounding boxes for anatomical structures, and AI diagnostic recommendations, compared with the traditional training group (c). Traditional training using the same 20 cases without AI assistance. (d) Learning retention evaluation using 200 fixed cases from OC-3000. (e) Generalization assessment using 100 random cases from OC-3000 to test diagnostic skill transferability. T-TG: traditional training groups; AI-TG: AI training groups; CL: cleft lip; CLP: cleft lip and palate.
Figure 3: Performance of the AIOC system on the OC-6000 testing dataset. (a) Confusion matrix, (b) ROC curve, and (c) t-SNE plot for case-based diagnosis (Control (blue), CL (orange), and CLP (green)). (d) Specific examples of Control, CL, and CLP cases, comparing the views, Gray-CAM results of AIOC, and the annotated key structures. Red boxes highlight cleft lip, cleft alveolus, or cleft palate abnormalities, while green boxes indicate normal tissue regions in control cases. CL: cleft lip; CLP: cleft lip and palate; CLV: cleft lip view; NLV: normal lip view; NAPV: normal alveolus and palate view; CAPV: cleft alveolus and palate view; AUC: area under the curve; ROC: receiver operating characteristic.
| Characteristic | AUC (%) | Sensitivity (%) | Specificity (%) | Accuracy | FNR | FPR | F1 | Youden | |
| (95% CI) | (95% CI) | (95% CI) | (%) | (%) | (%) | (%) | Index | ||
| OC-6000 | |||||||||
| Control | 98.25 | 99.60 | 96.70 | 99.15 | 0.40 | 3.30 | 99.50 | 0.96 | |
| (96.05 - 99.80) | (98.55 - 99.95) | (90.67 - 99.31) | |||||||
| CL | 90.72 | 85.71 | 99.65 | 99.15 | 14.29 | 0.35 | 87.80 | 0.85 | |
| (84.44 - 99.82) | (70.75 - 100.00) | (99.16 - 100.00) |
Table 1: Diagnostic performance of AIOC on the OC-6000, OC-GT3000, and OC-Early datasets, and comparison with senior, junior, and AIOC-assisted junior (Junior-AIOC) radiologists on the OC-GT3000 dataset. The OC-6000 presents AIOC’s fetal OC case-based diagnostic performance across Control, CL, and CLP categories. The ”AIOC” rows report the average performance metrics across these categories. The OC-GT3000 compares AIOC with Senior, Junior, and Junior-AIOC radiologists. They represent the majority voting across three senior, junior, and Junior-AIOC radiologists, respectively. The OC-Early shows AIOC’s performance in the early stage (14-17 weeks gestational age). AUC: area under the receiver operating characteristic curve; CI: confidence interval; FPR: false positive rate; FNR: false negative rate; F1: F1-score.
Figure 4: Fetal case-based diagnosis results of the AIOC, senior&junior radiologists, and junior radiologists assisted by AIOC (Junior-AIOC) on the OC-GT3000 dataset. (a) Confusion matrix and (b) ROC curve for AIOC, Senior, Junior, and Junior-AIOC diagnosis. (c) Bar plot showing F1 scores (18-28 weeks) for AIOC and radiologists (Senior&Junior). Due to the small number of cases in the 18-20 and 25-28 week groups, they were combined for analysis. (d) Representative examples of CLP, CL, and Control cases and their diagnosed results by AIOC, junior radiologists, senior radiologists, and junior radiologists assisted by AIOC. Each case includes multiple ultrasound views (left) and their zoom-in views of the region of interest (right). Red boxes highlight cleft lip, cleft alveolus, or cleft palate abnormalities, while green boxes indicate normal tissue regions in control cases. CL: cleft lip; CLP: cleft lip and palate; AUC: area under the curve; ROC: receiver operating characteristic.
Figure 5: Fetal case-based diagnosis results from the AIOC system on the OC-Early dataset. (a) Confusion matrix for fetal case-based diagnosis. (b) Specific cases: A CL case at 17 weeks of gestational age, with AIOC providing the correct diagnosis and Gray-CAM visualizing the diagnostic details. A CLP case at 15 weeks of gestational age, with AIOC also providing the correct diagnosis and visualizing the details. CL: cleft lip; CLP: cleft lip and palate.
Figure 6: Medical training results of the pilot trial involving radiologists and trainees in fetal OC and healthy case-based diagnosis. (a) Detailed quantitative analysis of average sensitivity (%) and box plots of trainees and radiologists in traditional training group (T-TG) and AI-augmented training group (AI-TG) during the four-cycle examination of 200 fixed cases. Box plots show the sensitivity of individual participants (n = 6 participants per box, represented by gray dots within each box). Box plots display median (center line), 25th–75th percentiles (box bounds), whiskers extending to the minimum and maximum values within 1.5 × IQR from the quartiles. (b) Learning curve of three specific cases (control, CL, and CLP) from the 200 fixed cases, diagnosed by one radiologist (T-TG-2-A, blue) from the T-TG group and one radiologist (AI-TG-2-A, green) from the AI-TG group. (c) Detailed quantitative analysis of average sensitivity (%) and bar plots of trainees and radiologists in two groups (T-TG and AI-TG) during the four-cycle examination of 100 random cases. Data are presented as the mean (measure of center) and 95% confidence intervals represented by error bars. (d) Results from four challenging cases across four examinations. The left panels show representative ultrasound views of each case, with anatomical landmarks annotated. Green boxes indicate normal structures, red boxes and yellow stars mark abnormal structures and cleft locations, respectively. The right panels display diagnostic outcomes for six participants (A–F) in each group: blue indicates the correct diagnosis for the T-TG group, green for the AI-TG group, and grey indicates an incorrect diagnosis. CL: cleft lip; CLP: cleft lip and palate; CLV: cleft lip view; NLV: normal lip view; NAPV: normal alveolus and palate view; CAPV: cleft alveolus and palate view; T-TG: traditional training groups; AI-TG: AI training groups.
Figure E1: View classification visualization of the AIOC system on the OC-6000 and OC-GT3000 datasets. a and d: Confusion matrix of the view classification of the OC-6000 and OC-GT3000 datasets; b and e: Receiver operating characteristic (ROC) curve of the OC-6000 and OC-GT3000 datasets; c and f: t-distributed Stochastic Neighbor Embedding (t-SNE) of the OC-6000 and OC-GT3000 datasets. CLV: cleft lip view; NLV: normal lip view; NAPV: normal alveolus and palate view; CAPV: cleft alveolus and palate view; AUC: area under the curve; ROC: receiver operating characteristic.
| Characteristic | AUC (%) | Sensitivity (%) | Specificity (%) | Accuracy | FNR | FPR | F1 | Youden |
| (95%CI) | (95%CI) | (95%CI) | (%) | (%) | (%) | (%) | Index | |
| Senior R1 | ||||||||
| Control | 98.7(97.75-99.73) | 99.87(97.31-100.00) | 97.87(95.81-99.94) | 99.87 | 0.13 | 2.13 | 99.87 | 0.98 |
| CL | 94.37(86.62-99.94) | 86.89(74.37-100.00) | 99.84(99.79-100.00) | 76.19 | 11.11 | 0.16 | 82.05 | 0.89 |
| CLP | 98.19(90.92-99.95) | 99.89(97.31-100.00) | 99.9(99.79-100.00) | 98.2 | 3.53 | 0.1 | 77.33 | 0.96 |
| Average | 97.11(94.37-99.69) | 95.50(89.27-100.00) | 99.2(98.43-99.97) | 91.42 | 4.92 | 0.8 | 93.08 | 0.94 |
Table E1: Diagnostic performance of senior radiologists, junior radiologists, and AIOC-assisted junior radiologists across the three types of diagnoses: control, cleft lip (CL), and cleft lip and palate (CLP), in the OC-GT3000 dataset. AUC: area under the receiver operating characteristic curve; CI: confidence interval; FPR: false positive rate; FNR: false negative rate; F1: F1-score.
| GROUP | Sensitivity | 95%CI | Specificity | 95%CI | Accuracy | 95%CI | -value |
| 200 fixed cases | |||||||
| Examination-1 | |||||||
| T-TG-1 | 55.19 | 44.67 - 66.25 | 84.15 | 78.5 - 89.76 | 83.22 | 78.19 - 88.25 | |
| AI-TG-1 | 67.55 | 64.41 - 90.99 | 89.98 | 90.23 - 97.36 | 88.17 | 88.51 - 95.71 | 0.17 |
| T-TG-2 | 77.49 | 52.10 - 82.93 | 93.86 | 85.08 - 94.79 | 92.11 | 83.88 - 92.45 | |
| AI-TG-2 | 85.86 | 74.46 - 95.66 | 95.00 | 91.82 - 98.00 | 94.11 | 91.0 - 97.14 | 0.60 |
Table E2: The Diagnostic performance between Traditional training group (T-TG) and AI-augmented training group (AI-TG) in the pilot trial of 200 fixed cases and 100 random cases; The difference in sensitivity between T-TG and AI-TG was assessed using a two-sided Mann–Whitney U test. To control for family-wise error rate, -values were adjusted for multiple comparisons using the Šidák correction. T-TG-1 and AI-TG-1: trainees; T-TG-2 and AI-TG-2: radiologists; CI: confidence interval.
Figure E2: Confusion matrix of senior radiologists, junior radiologists, and AIOC-assisted junior radiologists for the classification of three types of orofacial clefts: control, cleft lip (CL), and cleft lip and palate (CLP). a-c. Performance of senior R1, R2 and R3. d-f. Performance of junior R4, R5 and R6. g-i. Performance of AIOC-assisted junior radiologists R4, R5 and R6. CL: cleft lip; CLP: cleft lip and palate.
| Hospital | OC cases | healthy cases |
| Changshu Hospital of Traditional Chinese Medicine | 51 | 1601 |
| Changzhou Maternal and Child Health Care Hospital | 41 | 285 |
| Hangzhou Women’s Hospital | 100 | 0 |
| Jiangsu Taizhou People’s Hospital | 82 | 156 |
| Nanjing Women and Children’s Healthcare Hospital | 52 | 0 |
| Northwest Women’s and Children’s Hospital | 183 | 907 |
Table S1: List of 22 hospitals from our multicenter study and the amount of data provided for fetal OC and healthy cases. The total data includes both the included experimental data and the excluded data.
| Characteristic | AUC (%) | Sensitivity (%) | Specificity (%) | Accuracy | FNR | FPR | F1 | Youden |
| (95%CI) | (95%CI) | (95%CI) | (%) | (%) | (%) | (%) | Index | |
| OC-6000 | ||||||||
| NAPV | 99.49 | 98.2 | 99.92 | 99.27 | 1.8 | 0.51 | 97.9 | 0.98 |
| (99.85 - 99.98) | (97.03 - 99.36) | (99.21 - 99.78) | ||||||
| NLV | 98.94 | 99.13 | 99.84 | 99.06 | 0.87 | 1.06 | 99.26 | 0.98 |
| (99.67 - 99.95) | (98.7 - 99.55) | (98.32 - 99.56) |
Table S2: View classification performance of the AIOC system on the OC-6000 and OC-GT3000 datasets was evaluated across four key views: Normal Lip View (NLV), Normal Alveolus and Palate View (NAPV), Cleft Lip View (CLV), and Cleft Alveolus and Palate View (CAPV). Data are presented as %. AUC: area under the receiver operating characteristic curve; CI: confidence interval; FPR: false positive rate; FNR: false negative rate; F1: F1-score.
| Structure | upper lip | alveolar ridge | cleft lip | cleft alveolus | cleft palate |
| mAP of OC-6000 | 0.93 | 0.73 | 0.86 | 0.84 | 0.44 |
| mAP of OC-GT3000 | 0.99 | 0.97 | 0.80 | 0.77 | 0.42 |
Table S3: Structure detection performance of the AIOC system on the OC-6000 and OC-GT3000 datasets. mAP: mean Average Precision.
| GA/ Groups | 18 - 20 weeks | 21 weeks | 22 weeks | 23 weeks | 24 weeks | 25 - 28 weeks | SD |
| AIOC | 100.00 | 99.80 | 82.40 | 93.91 | 94.32 | 93.96 | 5.84 |
| Seniors | 82.22 | 99.93 | 100 | 95.79 | 92.99 | 100.00 | 6.35 |
| Juniors | 59.37 | 99.93 | 82.47 | 85.31 | 84.95 | 97.04 | 13.11 |
Table S4: Comparative performance based on F1-score for fetal OC case-based diagnosis across different gestational weeks (18–28 weeks), with AIOC compared to senior, junior, and junior-AIOC radiologists. Due to the small number of cases in the 18-20 and 25-28 week groups, these weeks were combined for statistical analysis. GA: gestational age; SD: standard deviation.
| Doctor | AI Accuracy | Doctor Accuracy | Automation Bias Metrics | ||
| Without AI | With AI | Over-reliance | Appropriate Reliance | ||
| Junior R4 | 99.5% | 99.6% | 99.8% | 11.8% | 0.4% |
| Junior R5 | 99.5% | 99.3% | 99.8% | 11.8% | 0.7% |
| Junior R6 | 99.5% | 98.5% | 99.4% | 5.9% | 1.2% |
Table S6: Automation Bias Analysis in AI-Assisted Diagnosis
| Type | Correct Num. | Correct Proportion | ||
| T-TG | AI-TG | |||
| Examination-1 | CL | 8 | 37.50% | 62.50% |
| Examination-2 | CLP | 7 | 28.57% | 71.43% |
| Examination-3 | CLP | 12 | 41.67% | 58.33% |
| Examination-4 | CL | 5 | 40.00% | 60.00% |
Table S7: Detailed results of four difficult cases out of 100 random cases in medical training, diagnosed by traditional (T-TG) and AI-augmented training groups (AI-TG).
Figure S1: Application of the AIOC system in orofacial case (OC) diagnosis and result display. (a) Web Interface – Includes interfaces for example cases and uploading new cases, predicting views and structures of each OC case image, and the final diagnosis prediction interface. (b) Mobile Application – Displays the initial case upload screen and interfaces for predicting views, structures, and diagnostic outcomes of OC cases. (c) Fetal OC and Healthy Cases – The workflow for using the AIOC system includes radiologists uploading OC and healthy cases to the AI interaction interface on the website or mobile application. Through joint decision-making with AIOC, the system supports diagnostic assistance and training purposes.
Figure S2: Grad-CAM visualizations for both correctly and incorrectly diagnosed cases. Each row displays two different ultrasound views per case. For each view, the left image shows Grad-CAM heatmap visualization highlighting regions of model attention (warmer colors indicate higher attention), and the right image shows the corresponding ultrasound with bounding boxes overlaid on anatomical structures. Top panel (Correct Diagnosis): Two cases where the model correctly identified CL and CLP. Bottom panel (Incorrect Diagnosis): Two cases demonstrating diagnostic errors, where ground truth labels differ from model predictions (GT: Control vs AIOC: CL in the first case; GT: CLP vs AIOC: CL in the second case).