Background Mid-trimester transabdominal ultrasound is the standard method for routine prenatal screening of brain anomalies. However, ultrasound is highly operator-dependent, leading to variability in detection rates. Deep learning (DL) models hold potential for improving the detection rates of fetal brain anomalies in mid-trimester ultrasounds. Objectives The primary objective of this research is to develop a DL-based model to classify fetal brain images from mid-trimester ultrasounds as either normal or pathological. Material and Methods Retrospective multicenter cohort study involving eight fetal medicine centers worldwide, which provided fetal brain images in a 2:1 ratio of normal to abnormal cases. We included 319 cases from patients undergoing prenatal screening during the second trimester, with 218 normal cases and 101 pathological cases. Each case comprised one transventricular and one transcerebellar image. Abnormal cases were confirmed by neonatal ultrasound, follow-up examination, or autopsy. All images underwent preprocessing and data augmentation before training. We developed two DL models: 1) a “Full Image Model”, developed to analyse the entire brain image and 2) a “6 Regions Model”, that focuses on six manually annotated regions. Both models were trained and validated using two approaches: a hold-out validation with 65% of data for training, 20% for validation, and 15% for testing, and a 3-fold cross-validation approach using 85% for training/validation and 15% for testing. Performance was evaluated with accuracy, sensitivity, specificity, F1-score, and AUC. Results In the hold-out validation, the “Full Image Model” failed to converge, with unstable performance metrics despite more than 100 training cycles. Consequently, we discarded this model before the testing phase. The “6 Regions Model” achieved an AUC of 0.983 (95% CI 0.944 – 1.000), an accuracy of 93.6 (95% CI 87.2 – 100.0), a sensitivity of 93.3% (95% CI 78.6 – 100.0), a specificity of 93.8% (95% CI 84.4 – 100.0), and a F1-score of 0.903 (95% CI 0.765 – 1.000) on the test dataset. The average classification time for the “6 Regions Model” was approximately 50 milliseconds per image. Conclusion This thesis presents a novel DL model for classifying fetal brain images from routine mid-trimester ultrasound scans as normal or pathological. The model demonstrated excellent performance metrics, highlighting its potential as a valuable tool in clinical and educational settings. Future advancements in model complexity and clinical workflows are necessary to understand the model’s full potential in global prenatal screening, aiming to improve maternal-fetal health outcomes worldwide.
Development of a Deep Learning algorithm to recognize abnormal findings at routine second trimester fetal brain ultrasound
Ruben, Ramirez Zegarra
2025
Abstract
Background Mid-trimester transabdominal ultrasound is the standard method for routine prenatal screening of brain anomalies. However, ultrasound is highly operator-dependent, leading to variability in detection rates. Deep learning (DL) models hold potential for improving the detection rates of fetal brain anomalies in mid-trimester ultrasounds. Objectives The primary objective of this research is to develop a DL-based model to classify fetal brain images from mid-trimester ultrasounds as either normal or pathological. Material and Methods Retrospective multicenter cohort study involving eight fetal medicine centers worldwide, which provided fetal brain images in a 2:1 ratio of normal to abnormal cases. We included 319 cases from patients undergoing prenatal screening during the second trimester, with 218 normal cases and 101 pathological cases. Each case comprised one transventricular and one transcerebellar image. Abnormal cases were confirmed by neonatal ultrasound, follow-up examination, or autopsy. All images underwent preprocessing and data augmentation before training. We developed two DL models: 1) a “Full Image Model”, developed to analyse the entire brain image and 2) a “6 Regions Model”, that focuses on six manually annotated regions. Both models were trained and validated using two approaches: a hold-out validation with 65% of data for training, 20% for validation, and 15% for testing, and a 3-fold cross-validation approach using 85% for training/validation and 15% for testing. Performance was evaluated with accuracy, sensitivity, specificity, F1-score, and AUC. Results In the hold-out validation, the “Full Image Model” failed to converge, with unstable performance metrics despite more than 100 training cycles. Consequently, we discarded this model before the testing phase. The “6 Regions Model” achieved an AUC of 0.983 (95% CI 0.944 – 1.000), an accuracy of 93.6 (95% CI 87.2 – 100.0), a sensitivity of 93.3% (95% CI 78.6 – 100.0), a specificity of 93.8% (95% CI 84.4 – 100.0), and a F1-score of 0.903 (95% CI 0.765 – 1.000) on the test dataset. The average classification time for the “6 Regions Model” was approximately 50 milliseconds per image. Conclusion This thesis presents a novel DL model for classifying fetal brain images from routine mid-trimester ultrasound scans as normal or pathological. The model demonstrated excellent performance metrics, highlighting its potential as a valuable tool in clinical and educational settings. Future advancements in model complexity and clinical workflows are necessary to understand the model’s full potential in global prenatal screening, aiming to improve maternal-fetal health outcomes worldwide.File | Dimensione | Formato | |
---|---|---|---|
PhD Thesis.pdf
embargo fino al 01/04/2026
Dimensione
1.63 MB
Formato
Adobe PDF
|
1.63 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/213319
URN:NBN:IT:UNIPR-213319