Numerous research works were done to remove speckle noise in ultrasound images. The noise reduction filters used were Lee, Frost filters to recent deep learning techniques for speckle noise reduction. Speckle Reduction Anisotropic Diffusion (SRAD) was highly preferred in the speckle noise reduction in recent years [Hyunho Choi and Jechang Jeong, 2020, I. Z. Nishu, et al., 2019]. Recent research works were primarily based on deep learning. Muhammad Shahin Uddin et al., 2016 used a complex wavelet-diffusion algorithm for noise reduction and preservation of edges in 2D and 3D ultrasound images. Rayleigh model was used for 2D images and a mixture of Maxwell model for 3D images. A genetic algorithm optimized the parameters. W.Zuo et al., 2018 used a Feedforward CNN (DnCNN) for denoising. The resultant noisy images were compared with the ground truth images.
Hyunho Choi and Jechang Jeong 2020 proposed a new hybrid model for speckle noise removal. SRAD, along with Discrete Wavelet Transform (DWT), image filtering using weight guiding technique, and gradient-domain image filtering, removed noise. By using the proposed method, the edges were preserved in a better way. The model's performance was compared with the conventional methods based on PSNR, SSIM, and MSE values. The algorithm was applied on ultrasound, airplane, Lena, boat, camera-man, and pepper images. Hamid Reza Shahdoosti, and Zahra Rahimi, 2019 used the Canny algorithm to extract edge maps. Initially, Shearlet transform was used to decompose the image into low-frequency components. Table 1 summarizes pre-processing techniques and the evaluation metrics of the filters.
Acoustic non-invasive data was considered [MD Erfanul Alam et al., 2019]. Support Vector Machine (SVM) was used to classify dysplasia into four classes. They were tested based upon the models that were stimulated by acoustic noise. The features extracted were a phase, transfer function, and coherence. The overall accuracy achieved was 79%, with the area under the curve as 0.93. Linear, Quadratic, and Linear SVM were considered for the study. Semiautomatic classification of developmental dysplasia was identified from 3D ultrasound images. The features extracted were skewness, kurtosis, convexity, alpha angle, and acetabular contact angle. A random forest classifier was used to classify acetabular shapes based on the geometric features. Three classifications, namely, normal, borderline, and dysplastic hips, were taken in the study, and the true positive rates achieved were 94.4%, 62.5%, and 89.9%, respectively [Hareendranathan, A. R, et al., 2016].
Table 1
Summary of pre-processing techniques
Author | Filters | Evaluation Metrics |
Muhammad Shahin Uddin et al., 2016 | Non-linear multi-scale complex wavelet diffusion. GA-EM model | SSIM, EdgeMSE, CNR, FoM, UIQI |
Chen et al., 2019 | Super-pixel segmentation of bilateral filtering and detailed compensation. | PSNR, SSIM |
Hyunho Choi and Jechang Jeong,2020 | SRAD, Discrete Wavelet Transform, Weighted Guided Image Filtering, Gradient Domain Guided Image Filtering | PSNR, SSIM, SMPI, Computational Cost in seconds. |
I.Z.Nishu et al., 2019 | SRAD, DWT | PSNR, RMSE, SSIM, Computational time in seconds. |
Yong Xu et al., 2015 | DNN, Non-linear regression function. | SNR, logMMSE |
Wang et al., 2013 | Modified PM with direct laplacian | PSNR, MSSIM |
Kumwilaisak et al., 2020 | Deep CNN and Multidirectional Long-Short Term Memory Architectures. | PSNR, SSIM |
Zuo et al., 2018 | Deep CNN, Single CNN | PSNR |
Shahdoosti et al., 2019 | Deep CNN, Canny algorithm, Non-Subsampled Shearlet Transform | PSNR |
Gai et al.,2019 | MP-DCNN, Residual learning | MSE, PSNR, SSIM |
With the aid of Artificial Intelligence (AI), a model was proposed that evaluated anteroposterior pelvic radiograph images. A total of 101 images were considered [Li, Qiang, et al., 2019]. Based on four key points, the Regional Convolutional Neural Network (R-CNN) classified the images. As observed, the diagnostic sensitivity and accuracy were close to those of physicians. It was possible to identify sharp angles in the image using the proposed method. The work in this paper used a standard mask R CNN, and it was fine tuned. 58 patients had Magnetic Resonance Imaging (MRI) images taken to detect focal cortical dysplasia lesions. Three machine learning algorithms were used to classify dysplasia based on morphological and intensity-based features [Ganji, Zohreh, et al., 2021]. The sensitivity, specificity, and accuracy of the decision tree were 96.7%, 100%, and 98.6%, SVM was 95%, 100%, and 97.9%, and for Artificial Neural Network (ANN) was 96.7%, 100%, and 98.6% respectively. Developmental dysplasia was identified and diagnosed considering 2601 images obtained from subjects below twelve years old. Using the template matching technique, the right and left side of the hip was matched. CNN was employed to detect the area of the sample image from the original image [Park, Hyoung, et al., 2021]. By utilizing this technique, significant results were obtained, and the results had a close resemblance to the radiologist's result.
A computational image analysis technique that automatically identified the images and extracts the metrics from 2D ultrasound images was identified [Quader N et al., 2017]. For this work, 54 ultrasound images of the femoral head of 10 infants were collected. The geometric features of bone/ cartilage boundaries were identified using the local phase symmetry-based method. A random forest classifier was used to classify the images based on the extracted features. Validation was done with 693 images acquired from 35 infants, and the area under the curve was 0.985. Hip joint ultrasound images were considered the gold standard for diagnosing dysplasia. Two-stage training of a deep-learning algorithm was proposed to identify the anatomical structures and mark the critical points in the images [Yann COADOU et al., 2013]. The R-CNN technique did the segmentation of the images. A free online tool named the Computer Vision Annotation Tool (CVAT) annotated videos and images. The image data was extracted in the first training, and the data marking was done in the second. In the final stage, the alpha and the beta angles were obtained, and 512 images were utilized for the study. Aysun Sezer and Hasan Basri Sezer, 2019 conducted a dysplasia study with 447 normal babies and 228 affected babies. Speckle noise reduction was made through a Bayesian non-local mean filter, and CNN was used for the classification. The CNN achieved an accuracy of 97.70%. Lee S-W et al., 2021 selected 320 images from 921 original images. Alpha and beta angles were identified from the image using a multi-point detection algorithm using Mask R-CNN.