TL;DR
This paper introduces a novel method for automatically recognizing and classifying standard fetal facial ultrasound planes (FFUSP) by fusing Local Binary Pattern (LBP) and Histogram of Oriented Gradient (HOG) texture features, then feeding them into a Support Vector Machine (SVM). This approach achieves a high accuracy of 94.67%, offering a more objective and efficient alternative to manual diagnosis, which is crucial for early detection of fetal abnormalities.
What Problem Does This Paper Address?
Prenatal ultrasound diagnosis is vital for screening fetal abnormalities, but accurately identifying standard fetal facial ultrasound planes (FFUSP) is a prerequisite for reliable diagnosis. Traditionally, this identification relies on manual screening by doctors, which is subjective, time-consuming, and prone to errors due to varying levels of expertise and fatigue. This variability can lead to misdiagnosis and a reduced efficiency in clinical practice. The paper aims to address this by developing an automated system that can accurately and efficiently recognize these standard planes.
Methods
The proposed method involves a three-step process:
- Texture Feature Extraction:
- Local Binary Pattern (LBP): This operator describes local texture features by comparing the gray value of a central pixel with its neighbors, creating a binary pattern. The paper uses a specific formula for LBP calculation, with parameters R (radius) and P (number of neighbors) explored.
- Histogram of Oriented Gradient (HOG): HOG is used to capture gradient information within an image. It calculates gradient amplitudes and directions for each pixel and then accumulates these into histograms for specific image regions (cells).
- Texture Feature Fusion: The extracted LBP and HOG feature vectors are concatenated to create a combined feature vector. This fusion aims to leverage the complementary information from both texture descriptors, enhancing the discriminative power of the features.
- Multiclassification Classifier: A Support Vector Machine (SVM) is employed for classification. Since SVM is primarily a binary classifier, a one-vs-rest strategy is used to create four binary classifiers to distinguish between the four classes: Ocular Axial Plane (OAP), Median Sagittal Plane (MSP), Nasolabial Coronal Plane (NCP), and Non-Standard Plane (N-SP).
The process also includes data preprocessing steps like image cropping and gray normalization to reduce noise and standardize images before feature extraction.
Datasets
The study utilized a dataset of 943 standard fetal facial ultrasound images and 350 non-standard plane images, totaling 1293 images. These images were collected from pregnant women between 20 and 24 weeks of gestation. The data was sourced from three Grade A hospitals and screened by senior sonographers to ensure accuracy. The dataset was divided into four categories:
- Ocular Axial Planes (OAP): 221 images
- Median Sagittal Planes (MSP): 298 images
- Nasolabial Coronal Planes (NCP): 424 images
- Non-Standard Planes (N-SP): 350 images
The data was subjected to five-fold cross-validation for experimental evaluation.
Key Insights and Ideas
- Fusion is Key: Combining LBP and HOG features significantly outperforms using either feature alone, highlighting the benefit of a multimodal approach to texture description for this task.
- SVM's Robustness: Despite the availability of deep learning methods, a traditional SVM classifier, when combined with effective feature engineering (LBP+HOG), demonstrates strong performance and is less complex to train.
- Non-Standard Planes Matter: Including non-standard planes, especially those visually similar to standard ones, significantly increases the difficulty of the classification task and provides a more realistic evaluation of the method's robustness.
- Parameter Sensitivity: While the method shows good stability across different parameters, the choice of cell size in HOG feature extraction and the (P, R) values for LBP can influence performance, though the impact is not always drastic.
Key Results
The proposed texture feature fusion method (LBP + HOG + SVM) achieved an overall recognition accuracy of 94.67%.
- The average precision was 94.27%.
- The average recall was 93.88%.
- The average F1-score was 94.08%.
Comparative experiments showed that the LBP+HOG fusion with SVM outperformed individual features (LBP or HOG alone) and also performed better than using KNN or Naive Bayes classifiers with the same fused features. The method demonstrated stability across different parameter settings, with accuracy consistently above 90%.
Limitations and Future Work
The authors acknowledge several limitations:
- Limited Scope: The method is primarily for rough classification and does not identify specific anatomical structures within the standard planes.
- Misclassification: Some ultrasound images are still misclassified.
- Healthy Fetus Data: The experimental data primarily consisted of healthy fetuses, which might simplify the recognition challenge compared to cases with abnormalities.
- Dataset Size: The dataset size is considered relatively small, limiting comparisons with deep learning models trained on larger datasets.
Future work includes:
- Establishing a more standardized and larger ultrasound image database.
- Developing methods to detect fundamental anatomical structures within the planes.
- Collecting data from pathological cases to improve robustness.
- Investigating techniques to overcome image differences caused by various ultrasonic instruments and scanning methods.
- Exploring more effective ways to predict similarity for improved quality control and clinical application.
Summary updated