PFP-LHCINCA: Pyramidal Fixed-Size Patch-Based Feature Extraction and Chi-Square Iterative Neighborhood Component...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Fetal Medicine

Background:

• Ultrasound (US) is crucial for fetal sex determination in pregnancies at risk of X-linked disorders or ambiguous genitalia.
• Misdiagnoses in fetal sex determination are common due to operator inexperience and technical imaging challenges.

Purpose of the Study:

• To develop an efficient, automated ultrasound-based fetal sex classification model.
• To enhance screening efficiency and minimize misclassification rates in fetal sex determination.

Main Methods:

• Developed the PFP-LHCINCA model using pyramidal fixed-size patch generation, local phase quantization (LPQ), and histogram of oriented gradients (HOG) for feature extraction.
• Employed Chi-square iterative neighborhood component analysis (CINCA) for feature selection, optimizing for k-nearest neighbor (kNN) misclassification rates.
• Trained and tested the model on 671 expert-labeled fetal US images (339 male, 332 female) using shallow classifiers (kNN, decision tree, Naïve Bayes, linear discriminant, SVM) with Bayesian hyperparameter optimization.

Main Results:

• The PFP-LHCINCA model demonstrated a fetal sex classification accuracy of ≥88% across five classifiers.
• Achieved superior accuracy rates exceeding 98% with kNN and SVM classifiers.
• Validated the model's performance on a substantial dataset of expert-annotated fetal US images.

Conclusions:

• Automated fetal sex classification using the PFP-LHCINCA model is feasible and highly accurate.
• The model shows promise for clinical application in fetal ultrasound screening for sex determination.
• The PFP-LHCINCA architecture is adaptable for deep learning models with larger datasets.