Semi-Supervised Fetal Anomaly Detection Using Deep Learning Based on Nuchal Translucency and Maternal Factors

Semi-Supervised Fetal Anomaly Detection Using Deep Learning Based on Nuchal Translucency and Maternal Factors | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Semi-Supervised Fetal Anomaly Detection Using Deep Learning Based on Nuchal Translucency and Maternal Factors Mostafa Kashani, Fatemeh Naghavi, Maryam Rostami This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8521430/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Early detection of fetal anomalies during prenatal screening is vital for reducing maternal-fetal complications. However, the scarcity of labeled datasets and the presence of "noisy" clinical labels pose significant challenges for traditional supervised methods. In this work, we propose a two-stage semi-supervised deep learning framework designed to minimize reliance on extensive labeling while enhancing diagnostic accuracy in small, real-world datasets. Method In this work, we propose a two-stage semi-supervised deep learning framework designed to minimize reliance on extensive labeling while enhancing diagnostic accuracy in small, real-world datasets. In the first stage, we employ an unsupervised Autoencoder to learn the inherent structure of normal data, generating statistical pseudo-labels for suspicious cases by calculating reconstruction errors. In the second stage, a supervised fully-connected classifier integrates these pseudo-labels with engineered features—such as the product of NT and maternal age, NT-to-fetal length/weight ratios, z-scores, and logarithmic NT transformations. By using the reconstruction error as a complementary feature, our model effectively classifies the fetal status into normal or anomalous. Our dataset, collected from 149 pregnant women between 11 to 13 weeks of gestation, underwent rigorous preprocessing, including KNN imputation and logarithmic scaling. Results Statistical analysis revealed significant differences (P < 0.001) in disability status, smoking habits, and combined NT features. Our proposed model achieved an accuracy of 95.6%, sensitivity of 97.5%, and an AUC of 0.977, with only one false negative and one false positive. These results demonstrate an ideal balance between maximizing high-risk case identification and reducing unnecessary costs. Conclusion This framework offers a novel, easy-to-implement template for intelligent prenatal screening in standard clinical settings. Fetal Anomaly Detection Semi-supervised Learning Autoencoder Nuchal Translucency (NT) Prenatal Screening Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction Prenatal screening stands as a cornerstone of maternal-fetal care, aimed at reducing neonatal mortality and improving long-term quality of life. Among various methods, Nuchal Translucency (NT) sonography in the first trimester is recognized as a critical non-invasive marker for assessing the risk of chromosomal abnormalities, such as Down syndrome, and major structural heart defects [ 1 ]. Despite its importance, accurately interpreting NT images alongside maternal factors—including age, BMI, and medical history—requires high levels of expertise. Many healthcare centers struggle with this process due to limited human resources and equipment [ 2 ]. Recent leaps in Artificial Intelligence (AI) and deep learning have opened new horizons for automating and refining this diagnostic process. Deep learning models can identify complex patterns in ultrasound images, often outperforming initial clinical assessments in screening for major congenital heart diseases (CHD) during the first trimester [ 3 – 5 ]. Rather than replacing specialists, these models act as a "second pair of eyes," increasing sensitivity while remaining generalizable to diverse community settings [ 6 ]. A primary hurdle in applying AI to fetal screening is the lack of high-quality labeled data and the "noise" often found in clinical labels. Protocols like PARADISE, which merge deep learning with statistical analysis for morphological anomalies, highlight the urgent need for intelligent decision-support systems [ 7 ]. Deep learning has already shown superior performance over traditional methods in detecting fetal intracranial anomalies, significantly improving detection rates in randomized controlled trials [ 8 ]. Beyond imaging, physiological signals such as cardiotocography (CTG) benefit from semi-supervised learning and domain adaptation, providing a clearer picture of fetal well-being [ 9 ]. These multimodal approaches allow for the simultaneous prediction of both common and rare genetic disorders [ 10 ]. Furthermore, automating fetal biometry and amniotic fluid volume measurements through end-to-end networks has reduced analysis time while boosting precision [ 11 ]. Systemic reviews in clinical engineering further confirm the vast potential of AI-assisted ultrasound in prenatal diagnostics [ 12 ]. Integrating maternal-fetal metadata with imaging is an emerging trend in the field. Custom convolutional networks can now predict standard fetal planes and enhance screening accuracy through information fusion [ 13 ]. In first-trimester cystic hygroma detection, deep learning-based ultrasound analysis has yielded promising outcomes [ 14 ]. Similarly, the LIFE protocol focuses on automated decision systems for early fetal echocardiography, positioning deep learning as a tool for more nuanced interpretation [ 15 ]. In the cardiac domain, machine learning algorithms are increasingly used to predict outcomes for fetal heart disease and detect four-chamber view anomalies [ 16 – 18 ]. Bimodal models that combine CTG and uterine contractions further strengthen health assessments during prenatal care [ 19 ]. Additionally, Fetal Anatomy Reasoning Networks (FARN) leverage global context and local topology to make anomaly detection more precise [ 20 ]. Despite these advancements, most existing research focuses heavily on image data, often overlooking tabular data such as NT measurements and maternal factors. We believe that combining these quantitative variables can create much stronger predictive indicators. Semi-supervised frameworks offer an ideal solution for medical contexts where labels are scarce. In this study, we develop a two-stage semi-supervised deep learning framework. We first use an unsupervised Autoencoder to learn the internal structure of "normal" data and generate pseudo-labels based on reconstruction errors. We then utilize a supervised classifier that incorporates these errors as an additional feature to detect anomalies. By focusing on real-world data from 149 mothers at 11–13 weeks of pregnancy, our approach not only improves screening accuracy but also remains highly feasible for clinical environments with limited resources. 2. Methodology In this research, we present a two-stage deep learning framework that integrates unsupervised and supervised approaches for fetal anomaly detection. The primary objective of this framework is to mitigate the dependency on limited and often "noisy" labeled data in the medical field, particularly in small-scale fetal datasets. The overall architecture of our proposed model is illustrated in Fig. 1 . 2.1 Data Preprocessing and Feature Engineering During the preprocessing phase, we focused on key maternal and fetal features, including Nuchal Translucency (NT), fetal weight and length, maternal age, and Body Mass Index (BMI). We first cleaned and standardized the data; missing values were handled using K-Nearest Neighbors (KNN) Imputation. To reduce the model's sensitivity to feature scales, we normalized all numerical features using the StandardScaler method. Furthermore, we developed a set of engineered features derived from the NT variable to enhance the discriminative power between normal and anomalous states. These include: Natural logarithm of NT (log-NT). Normalized z-scores. NT-to-fetal length and NT-to-fetal weight ratios. A composite feature calculated as the product of NT and maternal age (NT × Age). 2.2 Stage One: Unsupervised Representation Learning We initially designed an Autoencoder network to learn the internal structure of "normal" data without requiring labels. By learning the latent representation of the data, this model extracts fundamental physiological patterns from healthy samples. After training, we calculated the Reconstruction Error for each sample. Samples exhibiting an error above a specific threshold were flagged as suspicious. This process generates a set of Pseudo-labels that statistically reflect potential anomalies, thereby refining the accuracy of limited clinical labels. The Autoencoder architecture consists of an input layer, an encoding layer with a ReLU activation function, and a linear reconstruction output layer. 2.3 Stage Two: Supervised Classification As shown in Fig. 1 , these pseudo-labels were then used to train a neural network for classification. The classifier consists of several sequential fully-connected (dense) layers designed for feature extraction, culminating in an output layer for final status prediction. To prevent overfitting, we incorporated Dropout layers into the design. Given the inherent class imbalance in the dataset, we adjusted the class weights during training to maintain learning equilibrium. Crucially, the reconstruction error from the first stage was fed into the classifier as an additional input, allowing the model to directly account for the degree of deviation from the "normal" pattern during decision-making. 3. Results 3.1 Dataset Description To evaluate our method, we conducted experiments on real-world data collected from 149 pregnant women. Participants underwent NT sonography between 11 to 13 weeks of gestation at five health centers in West Ahvaz during 2021. Inclusion and Exclusion Criteria Inclusion required a gestational age of 11–13 weeks and NT sonography performed with standard equipment (e.g., GE Voluson E8). Exclusion criteria included incomplete data, lack of maternal consent, or scans performed outside the specified timeframe. Data were extracted from medical records, anonymized, and stored in a secure database. Feature Set Numerical variables included NT thickness (mm), fetal length (cm), fetal weight (g), maternal age (years), BMI, Gravida, Abortion count, preeclampsia status, Gestational Diabetes (GDM), smoking, addiction, and folic acid intake. Qualitative features included delivery type, maternal comorbidities (e.g., minor thalassemia, COVID-19, hypothyroidism), fetal sex, and anomaly type (e.g., Down syndrome or oligohydramnios). We applied a logarithmic transformation to the NT feature to improve its distribution and reduce skewness. Additionally, we utilized the interaction features (NT/Length, NT/Weight, and NT × Age) to reveal non-linear relationships. A binary variable was also assigned for the critical NT threshold: a value of 1 if NT > 3.5 mm, and 0 otherwise, based on standard clinical risk criteria. Finally, the data was split into 70% for training and 30% for testing. 3.2 Statistical Analysis We performed a comparative analysis between the normal and anomalous groups. The significant findings are summarized in Table 1 . Table 1 Statistical comparison of key features between Normal and Anomalous groups (Mean ± SD) Feature Normal Group (n = 141) Anomalous Group (n = 8) P-value Disability 1.16 ± 0.71 2.38 ± 2.07 < 0.001 Smoker 1.99 ± 0.12 1.75 ± 0.4 < 0.001 NT × Mother Age 52.78 ± 21.57 78.73 ± 23.18 0.004 NT / Length 0.04 ± 0.01 0.06 ± 0.02 0.004 NT (Absolute) 1.96 ± 0.61 2.89 ± 0.86 0.004 Preeclampsia 1.95 ± 0.22 1.75 ± 0.46 0.022 As observed in Table 1 , several key variables showed significant differences. The mean disability score in the anomalous group was markedly higher than in the normal group ( P < 0.001 ). The smoker variable also showed a significant difference ( P < 0.001 ), suggesting lifestyle impacts on fetal development. Furthermore, the interaction feature NT × Mother Age was significantly higher in the anomalous group (78.73 vs 52.78, P = 0.004 ), indicating that the synergy between maternal age and nuchal thickness is a potent risk indicator. Similarly, the NT/Length ratio was higher in anomalous cases ( P = 0.004 ), reflecting abnormal relative growth. The absolute NT thickness was also significantly greater in the anomalous group (2.89 mm vs 1.96 mm, P = 0.004 ), aligning with clinical evidence regarding its role as a marker for chromosomal disorders. In Fig. 2 a ) , we illustrate the distribution of NT thickness. At lower NT ranges (1.0–2.0 mm), normal samples are significantly more frequent, whereas anomalous cases increase as NT thickness rises, particularly above 2.5 mm. Figure 2 b ) provides a box plot analysis, showing that while there is some overlap, the median NT is higher in the anomalous group with a tighter Interquartile Range (IQR). 3.3 Model Performance Evaluation After training, our proposed model delivered excellent results on the test set. We achieved an overall accuracy of 95.6% , sensitivity (recall) of 97.5% , specificity of 80% , and a precision of 97.5% . The weighted F1-score was 97.5%. The Area Under the Curve (AUC) reached 0.977, as shown in Fig. 4 , indicating a superior ability to distinguish between normal and anomalous states. The confusion matrix (Fig. 3 ) revealed only one false negative and one false positive. From a clinical perspective, high sensitivity is paramount in screening to ensure high-risk cases are not missed. Our sensitivity of 97.5% confirms the model's effectiveness, while the high precision reduces the costs associated with unnecessary diagnostic follow-ups. 3.4 Implementation Details Our Autoencoder consisted of an input layer matched to the number of features and an 8-neuron encoding layer with ReLU activation. It was trained using Mean Squared Error (MSE) loss and the Adam optimizer for 200 epochs. For the classification stage, we utilized a fully-connected neural network with three hidden layers (64, 32, and 16 neurons respectively) using ReLU activation. The final output layer used a Sigmoid function for binary classification. We applied Dropout (rates of 0.3 and 0.2) to mitigate overfitting. The model was optimized using Adam with a learning rate of 0.001, ensuring stable convergence and high accuracy. 4. Discussion 4.1 Performance and Clinical Significance Our proposed framework achieved a remarkable accuracy of 95.6%, sensitivity of 97.5%, and an AUC of 0.977 on a test set derived from 149 real-world samples. These results, characterized by only one false negative and one false positive, hold substantial clinical significance. High sensitivity is the primary goal of prenatal screening to ensure maximum identification of high-risk cases, while a low false-positive rate minimizes unnecessary diagnostic costs and maternal anxiety [ 21 ]. By combining an unsupervised Autoencoder for pseudo-label generation with a supervised classifier, we effectively reduced the reliance on limited and noisy clinical labels. This approach allows the model to capture fundamental physiological patterns in normal data, which is crucial for medical datasets suffering from class imbalance. 4.2 Interpretation of Statistical Findings and Feature Engineering Statistical analysis highlighted significant differences ( P < 0.001 ) in features such as disability scores and smoking status, emphasizing the impact of maternal lifestyle on fetal development [ 22 ]. Engineered interaction features, such as NT × Maternal Age and the NT-to-fetal length ratio, revealed non-linear relationships that traditional models might overlook. These findings align with prior studies on predicting fetal complications like macrosomia [ 23 ]. Furthermore, the distribution of NT thickness (as seen in Figs. 2 ) confirms that the risk of anomaly increases significantly when NT exceeds 2.5 mm, reinforcing its role as a key chromosomal marker [ 24 ]. Our feature engineering process, including z-score normalization and log transformations, effectively corrected skewed distributions, mirroring successful machine learning approaches in predicting fetal growth restriction [ 24 ]. 4.3 Comparison with Supervised and Existing Methods Compared to purely supervised methods, our semi-supervised framework demonstrates clear superiority. The Autoencoder learns the inherent "normal" structure without labels, and the resulting pseudo-labels refine the accuracy of sparse clinical data. This is reminiscent of semi-supervised multi-task networks used for artifact detection in neonatal signals, which boost performance in data-limited scenarios [ 25 ]. Moreover, incorporating the reconstruction error as an input directly factors the deviation from normality into the final decision—a technique also observed in generative models for fetal brain structural anomaly detection [ 26 ]. Our AUC of 0.977 outperforms many imaging-based studies; for instance, while deep learning has achieved similar accuracy in renal anomaly detection, those models often lack the integration of NT and maternal metadata [ 27 ]. 4.4 Limitations and Future Directions From a clinical perspective, the balance of 97.5% sensitivity and 80% specificity is ideal for real-time screening [ 21 ]. These results resonate with AI applications in fetal cardiac anomalies, where high sensitivity is achieved in the 11–13 week window [ 28 ]. However, our relatively small sample size remains a limitation, highlighting the need for validation on larger, more diverse cohorts [ 22 ]. While we compensated for class imbalance using class weighting, larger studies might benefit from advanced techniques like SMOTE, as seen in ensemble learning for delivery mode prediction [ 29 ]. Our implementation, utilizing a compact Autoencoder and a three-layer classifier with Dropout, ensured stable convergence, comparable to MLP models for classifying ultrasound planes [ 30 ]. This tabular-focused framework has high potential for integration into existing sonographer workflows [ 31 ]. Looking ahead, we envision merging this approach with multimodal monitoring (fetal-maternal [ 32 ]) and advanced imaging like fetal MRI [ 33 ] to further strengthen non-invasive diagnostics [ 34 , 35 ]. 5. Conclusion In this study, we introduced a two-stage semi-supervised deep learning framework that represents a significant step toward accurate and early fetal anomaly detection. By intelligently combining an unsupervised Autoencoder to uncover hidden patterns in healthy data with a supervised classifier that utilizes statistical pseudo-labels, we minimized the need for large labeled datasets. Our model achieved an accuracy of 95.6% and an AUC of 0.977 in a real-world cohort of 149 pregnant women. With only two errors in the entire test set, this performance strikes an ideal balance between maximizing the identification of high-risk cases and preventing false alarms—a feat that can save lives and reduce healthcare costs in prenatal care. Ultimately, by relying on accessible features such as nuchal translucency thickness, maternal age, and lifestyle factors, our method opens a new door to intelligent, cost-effective screening that can be implemented in standard medical centers. Where anomaly detection previously relied heavily on highly specialized expertise or advanced equipment, our semi-supervised model delivers remarkable clinical precision even with small and imbalanced datasets. This achievement not only promises a brighter future for mothers and their infants but also serves as a robust template for AI applications in medical fields where data is scarce—demonstrating that innovation, simplicity, and accuracy can work hand-in-hand to change lives. Declarations Acknowledgements The authors gratefully acknowledge Ahvaz Jundishapure Umiversiry Of Medical Sciences for providing the space, facilities, and invaluable encouragement that enabled the successful completion of this research. Their support and commitment to fostering an environment conducive to academic inquiry were essential throughout this project. Ethical Approval This study was approved by the Institutional Ethics Committee of Ahvaz Jundishapure Umiversiry Of Medical Sciences (Ethics Approval Code: IR.AJUMS.MEDICINE.REC.1402.037). All procedures were conducted in accordance with the ethical standards outlined in the Declaration of Helsinki. Consent to Participate Informed consent was obtained from all participants prior to their involvement in the study. Consent to Publish All participants provided written consent for the publication of data, images, and other related information in this manuscript. Data Availability Statement The datasets generated and analyzed during the current study are not publicly available but they are available from the corresponding author on reasonable request. Authors Contributions Fist Author conducted the experiments, collected the data, performed the data analysis and drafted the manuscript. Second Author conceptualized and designed the study. Third Author validated the data and provided medical advice. All authors contributed to the interpretation of the results, critically revised the manuscript, and approved the final version. Funding Not applicable. This research did not receive any external funding. Competing Interests Not applicable. The authors declare that they have no competing interests, financial or otherwise, related to the work reported in this manuscript. References Zhang L, et al. Development and validation of a deep learning model to screen for trisomy 21 during the first trimester from nuchal ultrasonographic images. JAMA Netw open. 2022;5(6):e2217854–2217854. Ramirez Zegarra R, Ghi T. Use of artificial intelligence and deep learning in fetal ultrasound imaging. Ultrasound Obstet Gynecol. 2023;62(2):185–94. Zelop CM, et al. Artificial Intelligence for the Detection of Fetal Ultrasound Findings Concerning for Major Congenital Heart Defects. Obstetrics & Gynecology; 2022. p. 101097. Athalye C, et al. Deep learning model for prenatal congenital heart disease (CHD) screening can be applied to retrospective imaging from the community setting, outperforming initial clinical detection in a well-annotated cohort. Ultrasound Obstet gynecology: official J Int Soc Ultrasound Obstet Gynecol. 2024;63(1):44. Athalye C, et al. OC19. 04: A deep learning model for prenatal congenital heart disease screening generalises to a community setting and outperforms clinical detection. Ultrasound Obstet Gynecol. 2023;62:45–45. Belciug S. Autonomous fetal morphology scan: deep learning + clustering merger–the second pair of eyes behind the doctor. BMC Med Inf Decis Mak. 2024;24(1):102. Belciug S, et al. Pattern Recognition and Anomaly Detection in fetal morphology using Deep Learning and Statistical learning (PARADISE): protocol for the development of an intelligent decision support system using fetal morphology ultrasound scan to detect fetal congenital anomaly detection. BMJ open. 2024;14(2):e077366. Lin M, et al. Deep learning system improved detection efficacy of fetal intracranial malformations in a randomized controlled trial. NPJ Digit Med. 2023;6(1):191. Li J, et al. Multicentric intelligent cardiotocography signal interpretation using deep semi-supervised domain adaptation via minimax entropy and domain invariance. Comput Methods Programs Biomed. 2024;249:108145. Tang J et al. A Deep-Learning-Based Method Can Detect Both Common and Rare Genetic Disorders in Fetal Ultrasound. Biomedicines 2023, 11, 1756. 2023. Slimani S, et al. Fetal biometry and amniotic fluid volume assessment end-to-end automation using Deep Learning. Nat Commun. 2023;14(1):7047. Sriraam N, et al. Ultrasound imaging based recognition of prenatal anomalies: a systematic clinical engineering review. Progress Biomedical Eng. 2024;6(2):023002. Rauf F, et al. Artificial intelligence assisted common maternal fetal planes prediction from ultrasound images based on information fusion of customized convolutional neural networks. Front Med. 2024;11:1486995. Walker MC, et al. Using deep-learning in fetal ultrasound analysis for diagnosis of cystic hygroma in the first trimester. PLoS ONE. 2022;17(6):e0269323. Ungureanu A, et al. Learning deep architectures for the interpretation of first-trimester fetal echocardiography (LIFE)-a study protocol for developing an automated intelligent decision support system for early fetal echocardiography. BMC Pregnancy Childbirth. 2023;23(1):20. Nurmaini S, et al. Deep learning for improving the effectiveness of routine prenatal screening for major congenital heart diseases. J Clin Med. 2022;11(21):6454. Nield L, et al. MACHINE LEARNING ALGORITHMS TO PREDICT OUTCOME OF FETAL CARDIAC DISEASE. Can J Diabetes. 2023;47(7):S16. Abi Habib P, et al. 95 Application of machine learning in the detection of four chamber abnormalities. Am J Obstet Gynecol. 2024;230(1):S70. Qiu T, et al. AI-augmented prenatal care: a dual-modal fetal health assessment system integrating cardiotocography and uterine contraction synergy. Front Physiol. 2025;16:1638788. Zhao L, et al. Farn: fetal anatomy reasoning network for detection with global context semantic and local topology relationship. IEEE J Biomedical Health Inf. 2024;28(8):4866–77. Drukker L. Real-time identification of fetal anomalies on ultrasound using artificial intelligence: what's next? Chichester, UK: John Wiley & Sons, Ltd.; 2022. pp. 285–7. Islam MN, et al. Machine learning to predict pregnancy outcomes: a systematic review, synthesizing framework and future research agenda. BMC Pregnancy Childbirth. 2022;22(1):348. Gu X et al. Machine learning approach for the prediction of macrosomia. Visual Computing for Industry, Biomedicine, and Art, 2024. 7(1): p. 22. Teng LY, et al. Interpreting the role of nuchal fold for fetal growth restriction prediction using machine learning. Sci Rep. 2022;12(1):3907. Hermans T, et al. A multi-task and multi-channel convolutional neural network for semi-supervised neonatal artefact detection. J Neural Eng. 2023;20(2):026013. You S et al. Conditional deep generative normative modeling for structural and developmental anomaly detection in the fetal brain. NeuroImage, 2025: p. 121442. Miguel OX, et al. Deep learning prediction of renal anomalies for prenatal ultrasound diagnosis. Sci Rep. 2024;14(1):9013. Liu X, et al. Applications of artificial intelligence-powered prenatal diagnosis for congenital heart disease. Front Cardiovasc Med. 2024;11:1345761. Ferreira I, et al. Ensemble learning for fetal ultrasound and maternal–fetal data to predict mode of delivery after labor induction. Sci Rep. 2024;14(1):15275. Krishna TB, Kokil P. Automated classification of common maternal fetal ultrasound planes using multi-layer perceptron with deep feature integration. Biomed Signal Process Control. 2023;86:105283. Drukker L, et al. Clinical workflow of sonographers performing fetal anomaly ultrasound scans: deep-learning‐based analysis. Ultrasound Obstet Gynecol. 2022;60(6):759–65. Georgieva A et al. Fetal-maternal monitoring in the age of artificial intelligence and computer-aided decision support: A multidisciplinary perspective . 2022, Frontiers Media SA. p. 1007799. Zhang W, et al. A joint brain extraction and image quality assessment framework for fetal brain MRI slices. NeuroImage. 2024;290:120560. Cohen SM, Chen M, Sun L. Advancements in prenatal diagnosis: from noninvasive prenatal tests to novel fetal imaging . 2025, Frontiers. p. 1682161. Wang Q, et al. Applications of artificial intelligence in early childhood health management: a systematic review from fetal to pediatric periods. Front Pead. 2025;13:1613150. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8521430","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":587598827,"identity":"abd6f95b-3526-47d6-8243-c7b014b5b496","order_by":0,"name":"Mostafa Kashani","email":"","orcid":"","institution":"Electrical and Computer Engineering Department,Semnan University ,Semnan,Iran","correspondingAuthor":false,"prefix":"","firstName":"Mostafa","middleName":"","lastName":"Kashani","suffix":""},{"id":587598828,"identity":"1ef47e06-436d-4905-bdb5-8133d717f9d0","order_by":1,"name":"Fatemeh Naghavi","email":"","orcid":"","institution":"Electrical and Computer Engineering Department,Semnan University ,Semnan,Iran","correspondingAuthor":false,"prefix":"","firstName":"Fatemeh","middleName":"","lastName":"Naghavi","suffix":""},{"id":587598829,"identity":"36020590-131e-441d-a87c-60ca262fa6b9","order_by":2,"name":"Maryam Rostami","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA60lEQVRIiWNgGAWjYBAC+wYwJcHAwN7AYAATPYBPi8EBBsYGsBaeA6RpAVmUQKTDDI63P3/wM8eCQXfm4wfFvDl1DPztBxgPV+DRYt9zxrCxd5sEg9ntNANj3m2HGSTOJDAcPINHi51EDmMD7zaJ+m23E0BagL64wcBwsAGPFmOJ9IeNf0G23Dz+AailjkGekBbDGQmGzbwgLTd4QLYwMxgQ0mJw5ozhbFmQljM5BYZztx3mMTyT2IBfy/H2Bx/fAt1jdvz4NgMgQ07u+OHDH/FpQQZsoKjkYYBFFDGA+QHRSkfBKBgFo2BEAQCVR1GFymVSwwAAAABJRU5ErkJggg==","orcid":"","institution":"Ahvaz Jundishapur University of Medical Sciences","correspondingAuthor":true,"prefix":"","firstName":"Maryam","middleName":"","lastName":"Rostami","suffix":""}],"badges":[],"createdAt":"2026-01-05 12:38:32","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8521430/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8521430/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102296859,"identity":"c9883abd-5cc2-4db6-9236-e7d0f06c6215","added_by":"auto","created_at":"2026-02-10 10:22:21","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":283341,"visible":true,"origin":"","legend":"\u003cp\u003eOverall architecture of the proposed two-stage framework.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8521430/v1/947c812affaac003ccc48199.png"},{"id":102215064,"identity":"5f341725-4f9f-4bb2-b941-015732cffcd8","added_by":"auto","created_at":"2026-02-09 12:48:19","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":95275,"visible":true,"origin":"","legend":"\u003cp\u003ea) Box plot of NT thickness distribution. b) Distribution of NT thickness in Normal vs. Anomalous groups.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8521430/v1/3bc8b3f79500d28e0917efc5.png"},{"id":102215067,"identity":"b05f1169-7113-4194-8f18-2d8800005f20","added_by":"auto","created_at":"2026-02-09 12:48:19","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":66634,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion Matrix of the proposed model.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8521430/v1/814e9d94c61c6dcc33e108a2.png"},{"id":102297111,"identity":"9c2d4ad5-1155-4e2f-a911-415baa255c19","added_by":"auto","created_at":"2026-02-10 10:25:53","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":144416,"visible":true,"origin":"","legend":"\u003cp\u003eROC Curve of the proposed model.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8521430/v1/107a0b7b8f66a380972f0842.png"},{"id":104781331,"identity":"dfb3111c-eff4-41a2-9b93-4c8563f568a0","added_by":"auto","created_at":"2026-03-17 07:55:26","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1387040,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8521430/v1/5e71ace5-3aa8-42d6-8f3f-0d9d8fe411ed.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Semi-Supervised Fetal Anomaly Detection Using Deep Learning Based on Nuchal Translucency and Maternal Factors","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003ePrenatal screening stands as a cornerstone of maternal-fetal care, aimed at reducing neonatal mortality and improving long-term quality of life. Among various methods, Nuchal Translucency (NT) sonography in the first trimester is recognized as a critical non-invasive marker for assessing the risk of chromosomal abnormalities, such as Down syndrome, and major structural heart defects [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Despite its importance, accurately interpreting NT images alongside maternal factors\u0026mdash;including age, BMI, and medical history\u0026mdash;requires high levels of expertise. Many healthcare centers struggle with this process due to limited human resources and equipment [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eRecent leaps in Artificial Intelligence (AI) and deep learning have opened new horizons for automating and refining this diagnostic process. Deep learning models can identify complex patterns in ultrasound images, often outperforming initial clinical assessments in screening for major congenital heart diseases (CHD) during the first trimester [\u003cspan additionalcitationids=\"CR4\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Rather than replacing specialists, these models act as a \"second pair of eyes,\" increasing sensitivity while remaining generalizable to diverse community settings [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eA primary hurdle in applying AI to fetal screening is the lack of high-quality labeled data and the \"noise\" often found in clinical labels. Protocols like PARADISE, which merge deep learning with statistical analysis for morphological anomalies, highlight the urgent need for intelligent decision-support systems [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Deep learning has already shown superior performance over traditional methods in detecting fetal intracranial anomalies, significantly improving detection rates in randomized controlled trials [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Beyond imaging, physiological signals such as cardiotocography (CTG) benefit from semi-supervised learning and domain adaptation, providing a clearer picture of fetal well-being [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. These multimodal approaches allow for the simultaneous prediction of both common and rare genetic disorders [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Furthermore, automating fetal biometry and amniotic fluid volume measurements through end-to-end networks has reduced analysis time while boosting precision [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Systemic reviews in clinical engineering further confirm the vast potential of AI-assisted ultrasound in prenatal diagnostics [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIntegrating maternal-fetal metadata with imaging is an emerging trend in the field. Custom convolutional networks can now predict standard fetal planes and enhance screening accuracy through information fusion [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. In first-trimester cystic hygroma detection, deep learning-based ultrasound analysis has yielded promising outcomes [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Similarly, the LIFE protocol focuses on automated decision systems for early fetal echocardiography, positioning deep learning as a tool for more nuanced interpretation [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. In the cardiac domain, machine learning algorithms are increasingly used to predict outcomes for fetal heart disease and detect four-chamber view anomalies [\u003cspan additionalcitationids=\"CR17\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Bimodal models that combine CTG and uterine contractions further strengthen health assessments during prenatal care [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Additionally, Fetal Anatomy Reasoning Networks (FARN) leverage global context and local topology to make anomaly detection more precise [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eDespite these advancements, most existing research focuses heavily on image data, often overlooking tabular data such as NT measurements and maternal factors. We believe that combining these quantitative variables can create much stronger predictive indicators. Semi-supervised frameworks offer an ideal solution for medical contexts where labels are scarce.\u003c/p\u003e \u003cp\u003eIn this study, we develop a two-stage semi-supervised deep learning framework. We first use an unsupervised Autoencoder to learn the internal structure of \"normal\" data and generate pseudo-labels based on reconstruction errors. We then utilize a supervised classifier that incorporates these errors as an additional feature to detect anomalies. By focusing on real-world data from 149 mothers at 11\u0026ndash;13 weeks of pregnancy, our approach not only improves screening accuracy but also remains highly feasible for clinical environments with limited resources.\u003c/p\u003e"},{"header":"2. Methodology","content":"\u003cp\u003eIn this research, we present a two-stage deep learning framework that integrates unsupervised and supervised approaches for fetal anomaly detection. The primary objective of this framework is to mitigate the dependency on limited and often \"noisy\" labeled data in the medical field, particularly in small-scale fetal datasets. The overall architecture of our proposed model is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Data Preprocessing and Feature Engineering\u003c/h2\u003e \u003cp\u003eDuring the preprocessing phase, we focused on key maternal and fetal features, including Nuchal Translucency (NT), fetal weight and length, maternal age, and Body Mass Index (BMI). We first cleaned and standardized the data; missing values were handled using K-Nearest Neighbors (KNN) Imputation. To reduce the model's sensitivity to feature scales, we normalized all numerical features using the \u003cspan fontcategory=\"NonProportional\" class=\"\" name=\"Emphasis\"\u003eStandardScaler\u003c/span\u003e method.\u003c/p\u003e \u003cp\u003eFurthermore, we developed a set of engineered features derived from the NT variable to enhance the discriminative power between normal and anomalous states. These include:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eNatural logarithm of NT (log-NT).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eNormalized z-scores.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eNT-to-fetal length and NT-to-fetal weight ratios.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eA composite feature calculated as the product of NT and maternal age (NT \u0026times; Age).\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Stage One: Unsupervised Representation Learning\u003c/h2\u003e \u003cp\u003eWe initially designed an Autoencoder network to learn the internal structure of \"normal\" data without requiring labels. By learning the latent representation of the data, this model extracts fundamental physiological patterns from healthy samples. After training, we calculated the \u003cb\u003eReconstruction Error\u003c/b\u003e for each sample. Samples exhibiting an error above a specific threshold were flagged as suspicious. This process generates a set of \u003cb\u003ePseudo-labels\u003c/b\u003e that statistically reflect potential anomalies, thereby refining the accuracy of limited clinical labels. The Autoencoder architecture consists of an input layer, an encoding layer with a ReLU activation function, and a linear reconstruction output layer.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Stage Two: Supervised Classification\u003c/h2\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, these pseudo-labels were then used to train a neural network for classification. The classifier consists of several sequential fully-connected (dense) layers designed for feature extraction, culminating in an output layer for final status prediction. To prevent overfitting, we incorporated Dropout layers into the design. Given the inherent class imbalance in the dataset, we adjusted the class weights during training to maintain learning equilibrium. Crucially, the reconstruction error from the first stage was fed into the classifier as an additional input, allowing the model to directly account for the degree of deviation from the \"normal\" pattern during decision-making.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Dataset Description\u003c/h2\u003e \u003cp\u003eTo evaluate our method, we conducted experiments on real-world data collected from 149 pregnant women. Participants underwent NT sonography between 11 to 13 weeks of gestation at five health centers in West Ahvaz during 2021.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eInclusion and Exclusion Criteria\u003c/strong\u003e \u003cp\u003eInclusion required a gestational age of 11\u0026ndash;13 weeks and NT sonography performed with standard equipment (e.g., GE Voluson E8). Exclusion criteria included incomplete data, lack of maternal consent, or scans performed outside the specified timeframe. Data were extracted from medical records, anonymized, and stored in a secure database.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eFeature Set\u003c/strong\u003e \u003cp\u003eNumerical variables included NT thickness (mm), fetal length (cm), fetal weight (g), maternal age (years), BMI, Gravida, Abortion count, preeclampsia status, Gestational Diabetes (GDM), smoking, addiction, and folic acid intake. Qualitative features included delivery type, maternal comorbidities (e.g., minor thalassemia, COVID-19, hypothyroidism), fetal sex, and anomaly type (e.g., Down syndrome or oligohydramnios).\u003c/p\u003e \u003c/p\u003e \u003cp\u003eWe applied a logarithmic transformation to the NT feature to improve its distribution and reduce skewness. Additionally, we utilized the interaction features (NT/Length, NT/Weight, and NT \u0026times; Age) to reveal non-linear relationships. A binary variable was also assigned for the critical NT threshold: a value of 1 if NT\u0026thinsp;\u0026gt;\u0026thinsp;3.5 mm, and 0 otherwise, based on standard clinical risk criteria. Finally, the data was split into 70% for training and 30% for testing.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Statistical Analysis\u003c/h2\u003e \u003cp\u003eWe performed a comparative analysis between the normal and anomalous groups. The significant findings are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eStatistical comparison of key features between Normal and Anomalous groups (Mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeature\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNormal Group (n\u0026thinsp;=\u0026thinsp;141)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAnomalous Group (n\u0026thinsp;=\u0026thinsp;8)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eP-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDisability\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e1.16\u0026thinsp;\u0026plusmn;\u0026thinsp;0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e2.38\u0026thinsp;\u0026plusmn;\u0026thinsp;2.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSmoker\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e1.99\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e1.75\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNT \u0026times; Mother Age\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e52.78\u0026thinsp;\u0026plusmn;\u0026thinsp;21.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e78.73\u0026thinsp;\u0026plusmn;\u0026thinsp;23.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.004\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNT / Length\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.04\u0026thinsp;\u0026plusmn;\u0026thinsp;0.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.06\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.004\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eNT (Absolute)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e1.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e2.89\u0026thinsp;\u0026plusmn;\u0026thinsp;0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.004\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePreeclampsia\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e1.95\u0026thinsp;\u0026plusmn;\u0026thinsp;0.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e1.75\u0026thinsp;\u0026plusmn;\u0026thinsp;0.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.022\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eAs observed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, several key variables showed significant differences. The mean \u003cb\u003edisability\u003c/b\u003e score in the anomalous group was markedly higher than in the normal group (\u003cem\u003eP\u0026thinsp;\u0026lt;\u0026thinsp;0.001\u003c/em\u003e). The \u003cb\u003esmoker\u003c/b\u003e variable also showed a significant difference (\u003cem\u003eP\u0026thinsp;\u0026lt;\u0026thinsp;0.001\u003c/em\u003e), suggesting lifestyle impacts on fetal development.\u003c/p\u003e \u003cp\u003eFurthermore, the interaction feature \u003cb\u003eNT \u0026times; Mother Age\u003c/b\u003e was significantly higher in the anomalous group (78.73 vs 52.78, \u003cem\u003eP\u0026thinsp;=\u0026thinsp;0.004\u003c/em\u003e), indicating that the synergy between maternal age and nuchal thickness is a potent risk indicator. Similarly, the \u003cb\u003eNT/Length\u003c/b\u003e ratio was higher in anomalous cases (\u003cem\u003eP\u0026thinsp;=\u0026thinsp;0.004\u003c/em\u003e), reflecting abnormal relative growth. The absolute \u003cb\u003eNT thickness\u003c/b\u003e was also significantly greater in the anomalous group (2.89 mm vs 1.96 mm, \u003cem\u003eP\u0026thinsp;=\u0026thinsp;0.004\u003c/em\u003e), aligning with clinical evidence regarding its role as a marker for chromosomal disorders.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea\u003cb\u003e)\u003c/b\u003e, we illustrate the distribution of NT thickness. At lower NT ranges (1.0\u0026ndash;2.0 mm), normal samples are significantly more frequent, whereas anomalous cases increase as NT thickness rises, particularly above 2.5 mm. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb\u003cb\u003e)\u003c/b\u003e provides a box plot analysis, showing that while there is some overlap, the median NT is higher in the anomalous group with a tighter Interquartile Range (IQR).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Model Performance Evaluation\u003c/h2\u003e \u003cp\u003eAfter training, our proposed model delivered excellent results on the test set. We achieved an overall \u003cb\u003eaccuracy of 95.6%\u003c/b\u003e, \u003cb\u003esensitivity (recall) of 97.5%\u003c/b\u003e, \u003cb\u003especificity of 80%\u003c/b\u003e, and a \u003cb\u003eprecision of 97.5%\u003c/b\u003e. The weighted \u003cb\u003eF1-score\u003c/b\u003e was 97.5%.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe \u003cb\u003eArea Under the Curve (AUC)\u003c/b\u003e reached 0.977, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, indicating a superior ability to distinguish between normal and anomalous states. The confusion matrix (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) revealed only one false negative and one false positive. From a clinical perspective, high sensitivity is paramount in screening to ensure high-risk cases are not missed. Our sensitivity of 97.5% confirms the model's effectiveness, while the high precision reduces the costs associated with unnecessary diagnostic follow-ups.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Implementation Details\u003c/h2\u003e \u003cp\u003eOur Autoencoder consisted of an input layer matched to the number of features and an 8-neuron encoding layer with ReLU activation. It was trained using Mean Squared Error (MSE) loss and the Adam optimizer for 200 epochs.\u003c/p\u003e \u003cp\u003eFor the classification stage, we utilized a fully-connected neural network with three hidden layers (64, 32, and 16 neurons respectively) using ReLU activation. The final output layer used a Sigmoid function for binary classification. We applied Dropout (rates of 0.3 and 0.2) to mitigate overfitting. The model was optimized using Adam with a learning rate of 0.001, ensuring stable convergence and high accuracy.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Performance and Clinical Significance\u003c/h2\u003e \u003cp\u003eOur proposed framework achieved a remarkable accuracy of 95.6%, sensitivity of 97.5%, and an AUC of 0.977 on a test set derived from 149 real-world samples. These results, characterized by only one false negative and one false positive, hold substantial clinical significance. High sensitivity is the primary goal of prenatal screening to ensure maximum identification of high-risk cases, while a low false-positive rate minimizes unnecessary diagnostic costs and maternal anxiety [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. By combining an unsupervised Autoencoder for pseudo-label generation with a supervised classifier, we effectively reduced the reliance on limited and noisy clinical labels. This approach allows the model to capture fundamental physiological patterns in normal data, which is crucial for medical datasets suffering from class imbalance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Interpretation of Statistical Findings and Feature Engineering\u003c/h2\u003e \u003cp\u003eStatistical analysis highlighted significant differences (\u003cem\u003eP\u0026thinsp;\u0026lt;\u0026thinsp;0.001\u003c/em\u003e) in features such as disability scores and smoking status, emphasizing the impact of maternal lifestyle on fetal development [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Engineered interaction features, such as NT \u0026times; Maternal Age and the NT-to-fetal length ratio, revealed non-linear relationships that traditional models might overlook. These findings align with prior studies on predicting fetal complications like macrosomia [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. Furthermore, the distribution of NT thickness (as seen in Figs.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) confirms that the risk of anomaly increases significantly when NT exceeds 2.5 mm, reinforcing its role as a key chromosomal marker [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Our feature engineering process, including z-score normalization and log transformations, effectively corrected skewed distributions, mirroring successful machine learning approaches in predicting fetal growth restriction [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Comparison with Supervised and Existing Methods\u003c/h2\u003e \u003cp\u003eCompared to purely supervised methods, our semi-supervised framework demonstrates clear superiority. The Autoencoder learns the inherent \"normal\" structure without labels, and the resulting pseudo-labels refine the accuracy of sparse clinical data. This is reminiscent of semi-supervised multi-task networks used for artifact detection in neonatal signals, which boost performance in data-limited scenarios [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Moreover, incorporating the reconstruction error as an input directly factors the deviation from normality into the final decision\u0026mdash;a technique also observed in generative models for fetal brain structural anomaly detection [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Our AUC of 0.977 outperforms many imaging-based studies; for instance, while deep learning has achieved similar accuracy in renal anomaly detection, those models often lack the integration of NT and maternal metadata [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Limitations and Future Directions\u003c/h2\u003e \u003cp\u003eFrom a clinical perspective, the balance of 97.5% sensitivity and 80% specificity is ideal for real-time screening [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. These results resonate with AI applications in fetal cardiac anomalies, where high sensitivity is achieved in the 11\u0026ndash;13 week window [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. However, our relatively small sample size remains a limitation, highlighting the need for validation on larger, more diverse cohorts [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. While we compensated for class imbalance using class weighting, larger studies might benefit from advanced techniques like SMOTE, as seen in ensemble learning for delivery mode prediction [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eOur implementation, utilizing a compact Autoencoder and a three-layer classifier with Dropout, ensured stable convergence, comparable to MLP models for classifying ultrasound planes [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. This tabular-focused framework has high potential for integration into existing sonographer workflows [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. Looking ahead, we envision merging this approach with multimodal monitoring (fetal-maternal [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]) and advanced imaging like fetal MRI [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e] to further strengthen non-invasive diagnostics [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e, \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eIn this study, we introduced a two-stage semi-supervised deep learning framework that represents a significant step toward accurate and early fetal anomaly detection. By intelligently combining an unsupervised Autoencoder to uncover hidden patterns in healthy data with a supervised classifier that utilizes statistical pseudo-labels, we minimized the need for large labeled datasets. Our model achieved an accuracy of 95.6% and an AUC of 0.977 in a real-world cohort of 149 pregnant women. With only two errors in the entire test set, this performance strikes an ideal balance between maximizing the identification of high-risk cases and preventing false alarms\u0026mdash;a feat that can save lives and reduce healthcare costs in prenatal care.\u003c/p\u003e \u003cp\u003eUltimately, by relying on accessible features such as nuchal translucency thickness, maternal age, and lifestyle factors, our method opens a new door to intelligent, cost-effective screening that can be implemented in standard medical centers. Where anomaly detection previously relied heavily on highly specialized expertise or advanced equipment, our semi-supervised model delivers remarkable clinical precision even with small and imbalanced datasets. This achievement not only promises a brighter future for mothers and their infants but also serves as a robust template for AI applications in medical fields where data is scarce\u0026mdash;demonstrating that innovation, simplicity, and accuracy can work hand-in-hand to change lives.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors gratefully acknowledge Ahvaz Jundishapure Umiversiry Of Medical Sciences \u0026nbsp;for providing the space, facilities, and invaluable encouragement that enabled the successful completion of this research. Their support and commitment to fostering an environment conducive to academic inquiry were essential throughout this project.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical Approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Institutional Ethics Committee of Ahvaz Jundishapure Umiversiry Of Medical Sciences \u0026nbsp;(Ethics Approval Code: IR.AJUMS.MEDICINE.REC.1402.037). All procedures were conducted in accordance with the ethical standards outlined in the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to Participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eInformed consent was obtained from all participants prior to their involvement in the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to Publish\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll participants provided written consent for the publication of data, images, and other related information in this manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets generated and analyzed during the current study are not publicly available but they are available from the corresponding author on reasonable request.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFist Author\u003c/strong\u003e conducted the experiments, collected the data, performed the data analysis and drafted the manuscript. \u003cstrong\u003eSecond Author\u003c/strong\u003e conceptualized and designed the study. \u003cstrong\u003eThird Author\u003c/strong\u003e validated the data and provided medical advice. All authors contributed to the interpretation of the results, critically revised the manuscript, and approved the final version.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable. This research did not receive any external funding.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable. The authors declare that they have no competing interests, financial or otherwise, related to the work reported in this manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eZhang L, et al. Development and validation of a deep learning model to screen for trisomy 21 during the first trimester from nuchal ultrasonographic images. JAMA Netw open. 2022;5(6):e2217854\u0026ndash;2217854.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRamirez Zegarra R, Ghi T. Use of artificial intelligence and deep learning in fetal ultrasound imaging. Ultrasound Obstet Gynecol. 2023;62(2):185\u0026ndash;94.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZelop CM, et al. Artificial Intelligence for the Detection of Fetal Ultrasound Findings Concerning for Major Congenital Heart Defects. Obstetrics \u0026amp; Gynecology; 2022. p. 101097.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAthalye C, et al. Deep learning model for prenatal congenital heart disease (CHD) screening can be applied to retrospective imaging from the community setting, outperforming initial clinical detection in a well-annotated cohort. Ultrasound Obstet gynecology: official J Int Soc Ultrasound Obstet Gynecol. 2024;63(1):44.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAthalye C, et al. OC19. 04: A deep learning model for prenatal congenital heart disease screening generalises to a community setting and outperforms clinical detection. Ultrasound Obstet Gynecol. 2023;62:45\u0026ndash;45.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBelciug S. Autonomous fetal morphology scan: deep learning\u0026thinsp;+\u0026thinsp;clustering merger\u0026ndash;the second pair of eyes behind the doctor. BMC Med Inf Decis Mak. 2024;24(1):102.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBelciug S, et al. Pattern Recognition and Anomaly Detection in fetal morphology using Deep Learning and Statistical learning (PARADISE): protocol for the development of an intelligent decision support system using fetal morphology ultrasound scan to detect fetal congenital anomaly detection. BMJ open. 2024;14(2):e077366.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin M, et al. Deep learning system improved detection efficacy of fetal intracranial malformations in a randomized controlled trial. NPJ Digit Med. 2023;6(1):191.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi J, et al. Multicentric intelligent cardiotocography signal interpretation using deep semi-supervised domain adaptation via minimax entropy and domain invariance. Comput Methods Programs Biomed. 2024;249:108145.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTang J et al. \u003cem\u003eA Deep-Learning-Based Method Can Detect Both Common and Rare Genetic Disorders in Fetal Ultrasound. Biomedicines\u003c/em\u003e 2023, 11, 1756. 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSlimani S, et al. Fetal biometry and amniotic fluid volume assessment end-to-end automation using Deep Learning. Nat Commun. 2023;14(1):7047.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSriraam N, et al. Ultrasound imaging based recognition of prenatal anomalies: a systematic clinical engineering review. Progress Biomedical Eng. 2024;6(2):023002.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRauf F, et al. Artificial intelligence assisted common maternal fetal planes prediction from ultrasound images based on information fusion of customized convolutional neural networks. Front Med. 2024;11:1486995.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWalker MC, et al. Using deep-learning in fetal ultrasound analysis for diagnosis of cystic hygroma in the first trimester. PLoS ONE. 2022;17(6):e0269323.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUngureanu A, et al. Learning deep architectures for the interpretation of first-trimester fetal echocardiography (LIFE)-a study protocol for developing an automated intelligent decision support system for early fetal echocardiography. BMC Pregnancy Childbirth. 2023;23(1):20.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNurmaini S, et al. Deep learning for improving the effectiveness of routine prenatal screening for major congenital heart diseases. J Clin Med. 2022;11(21):6454.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNield L, et al. MACHINE LEARNING ALGORITHMS TO PREDICT OUTCOME OF FETAL CARDIAC DISEASE. Can J Diabetes. 2023;47(7):S16.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbi Habib P, et al. 95 Application of machine learning in the detection of four chamber abnormalities. Am J Obstet Gynecol. 2024;230(1):S70.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQiu T, et al. AI-augmented prenatal care: a dual-modal fetal health assessment system integrating cardiotocography and uterine contraction synergy. Front Physiol. 2025;16:1638788.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao L, et al. Farn: fetal anatomy reasoning network for detection with global context semantic and local topology relationship. IEEE J Biomedical Health Inf. 2024;28(8):4866\u0026ndash;77.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDrukker L. Real-time identification of fetal anomalies on ultrasound using artificial intelligence: what's next? Chichester, UK: John Wiley \u0026amp; Sons, Ltd.; 2022. pp. 285\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIslam MN, et al. Machine learning to predict pregnancy outcomes: a systematic review, synthesizing framework and future research agenda. BMC Pregnancy Childbirth. 2022;22(1):348.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGu X et al. \u003cem\u003eMachine learning approach for the prediction of macrosomia.\u003c/em\u003e Visual Computing for Industry, Biomedicine, and Art, 2024. 7(1): p. 22.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTeng LY, et al. Interpreting the role of nuchal fold for fetal growth restriction prediction using machine learning. Sci Rep. 2022;12(1):3907.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHermans T, et al. A multi-task and multi-channel convolutional neural network for semi-supervised neonatal artefact detection. J Neural Eng. 2023;20(2):026013.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYou S et al. \u003cem\u003eConditional deep generative normative modeling for structural and developmental anomaly detection in the fetal brain.\u003c/em\u003e NeuroImage, 2025: p. 121442.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMiguel OX, et al. Deep learning prediction of renal anomalies for prenatal ultrasound diagnosis. Sci Rep. 2024;14(1):9013.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu X, et al. Applications of artificial intelligence-powered prenatal diagnosis for congenital heart disease. Front Cardiovasc Med. 2024;11:1345761.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFerreira I, et al. Ensemble learning for fetal ultrasound and maternal\u0026ndash;fetal data to predict mode of delivery after labor induction. Sci Rep. 2024;14(1):15275.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKrishna TB, Kokil P. Automated classification of common maternal fetal ultrasound planes using multi-layer perceptron with deep feature integration. Biomed Signal Process Control. 2023;86:105283.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDrukker L, et al. Clinical workflow of sonographers performing fetal anomaly ultrasound scans: deep-learning‐based analysis. Ultrasound Obstet Gynecol. 2022;60(6):759\u0026ndash;65.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGeorgieva A et al. \u003cem\u003eFetal-maternal monitoring in the age of artificial intelligence and computer-aided decision support: A multidisciplinary perspective\u003c/em\u003e. 2022, Frontiers Media SA. p. 1007799.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang W, et al. A joint brain extraction and image quality assessment framework for fetal brain MRI slices. NeuroImage. 2024;290:120560.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCohen SM, Chen M, Sun L. \u003cem\u003eAdvancements in prenatal diagnosis: from noninvasive prenatal tests to novel fetal imaging\u003c/em\u003e. 2025, Frontiers. p. 1682161.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Q, et al. Applications of artificial intelligence in early childhood health management: a systematic review from fetal to pediatric periods. Front Pead. 2025;13:1613150.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Fetal Anomaly Detection, Semi-supervised Learning, Autoencoder, Nuchal Translucency (NT), Prenatal Screening","lastPublishedDoi":"10.21203/rs.3.rs-8521430/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8521430/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eEarly detection of fetal anomalies during prenatal screening is vital for reducing maternal-fetal complications. However, the scarcity of labeled datasets and the presence of \"noisy\" clinical labels pose significant challenges for traditional supervised methods. In this work, we propose a two-stage semi-supervised deep learning framework designed to minimize reliance on extensive labeling while enhancing diagnostic accuracy in small, real-world datasets.\u003c/p\u003e\u003ch2\u003eMethod\u003c/h2\u003e \u003cp\u003eIn this work, we propose a two-stage semi-supervised deep learning framework designed to minimize reliance on extensive labeling while enhancing diagnostic accuracy in small, real-world datasets. In the first stage, we employ an unsupervised Autoencoder to learn the inherent structure of normal data, generating statistical pseudo-labels for suspicious cases by calculating reconstruction errors. In the second stage, a supervised fully-connected classifier integrates these pseudo-labels with engineered features\u0026mdash;such as the product of NT and maternal age, NT-to-fetal length/weight ratios, z-scores, and logarithmic NT transformations. By using the reconstruction error as a complementary feature, our model effectively classifies the fetal status into normal or anomalous. Our dataset, collected from 149 pregnant women between 11 to 13 weeks of gestation, underwent rigorous preprocessing, including KNN imputation and logarithmic scaling.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eStatistical analysis revealed significant differences (P\u0026thinsp;\u0026lt;\u0026thinsp;0.001) in disability status, smoking habits, and combined NT features. Our proposed model achieved an accuracy of 95.6%, sensitivity of 97.5%, and an AUC of 0.977, with only one false negative and one false positive. These results demonstrate an ideal balance between maximizing high-risk case identification and reducing unnecessary costs.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThis framework offers a novel, easy-to-implement template for intelligent prenatal screening in standard clinical settings.\u003c/p\u003e","manuscriptTitle":"Semi-Supervised Fetal Anomaly Detection Using Deep Learning Based on Nuchal Translucency and Maternal Factors","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-09 12:48:14","doi":"10.21203/rs.3.rs-8521430/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e3136e94-965c-4e7f-8289-dce5a8a65ec6","owner":[],"postedDate":"February 9th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-13T12:26:26+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-09 12:48:14","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8521430","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8521430","identity":"rs-8521430","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}