Nowadays, medical imaging plays a key role in the clinical workflows and has a capability of representing anatomical and physiological features that are otherwise inaccessible to inspection, thus proposing accurate imaging biomarkers and clinically useful information. The effectiveness of such techniques can be enhanced by the newly developed artificial intelligence (AI) methods. The new techniques have many applications in the computer-aided medical systems and clinical measurement.
Park et al. proposed an in-vitro flow model as a reference to analyze the characteristics of time-velocity curve acquired by phase-contrast magnetic resonance imaging (PC-MRI). Time-velocity curves of PC-MRI were compared with Doppler ultrasonography and electromagnetic flowmeter. Results showed that PC-MRI reported lower maximum velocity value and higher minimum velocity value than Doppler ultrasonography.
Automatic segmentation of prostate MRI faces big challenge since the low contrast of tissue boundary and the small effective area. Geng et al. proposed a novel deep learning-based network, consisting of an encoder-decoder structure with dense dilated spatial pyramid pooling. Experimental results showed a high accuracy and robustness for the proposed method, with a dice similarity coefficient of 0.954 and Hausdorff distance of 1.752 mm.
Wang et al. proposed a patient-specific probabilistic atlas-based method combing modified distance regularized level set for liver segmentation. Firstly, the similarities between training atlases and testing patient image were calculated to generate a series of weighted atlas, and were further used to generate the patient-specific probabilistic atlas. Then, a most likely liver region was determined based on the patient-specific probabilistic atlas. Finally, refinement was performed by the modified distance regularized level set model. The proposed scheme was evaluated on 35 public datasets, and the result verified the proposed method can give robust and precise liver segmentation.
Geng et al. developed a segmentation method of computed tomography (CT) image based on the combination of VGG-16 model and dilated convolution for lung cancer diagnosis. Experimental results over 137 images showed that this method can effectively segment the lung parenchymal area.
Chou et al. proposed a quantitative assessment method of artifact reduction techniques for CT imaging using a customized phantom. The phantom was scanned using three CT scanners (Toshiba, GE, Siemens) under the abdomen setting model. Fraction of affected pixel area obtained using single-energy metal artifact reduction of Toshiba scanner was 2.2% ± 0.7%, which was significantly lower than 4.1% ± 0.7% obtained using metal artifact reduction software of GE scanner. Single-energy metal artifact reduction technique in Toshiba scanner was recommended.
Liu et al. proposed a framework combining OneCut algorithm and adjacent image geometric features to continuously extract the region of interest (ROI) in targeted images. Firstly, image was segmented with the OneCut algorithm. Then, the foreground image (obtained ROI) was corroded into several seed points and the background image (other region except for ROI) was refined into a skeleton. Experimental results verified the reliable extraction of ROI.
Chen et al. introduced a non-linear Gaussian-form solution assumption-based imaging method derived from non-linear wave equation, which can be applied in the pulse-echo mode in diagnostic ultrasound. Herein the novelty is that, with two pulse transmission, only non-linear effects are reserved and other effects can be reduced.
Zhang et al. proposed a hybrid method consisting of random over sampling example, K-mean clustering and support vector machine (SVM), to overcome the imbalanced classification problem in diagnosis of breast cancer. Random over sampling example was firstly utilized to balance the dataset and diagnosis accuracy was further improved by K-mean clustering and SVM model. Experiment was implemented on breast cancer dataset and the other three datasets from the machine learning repository in University of California Irvine, and the results showed the new hybrid method outperformed the competitive algorithms in term of G-mean and accuracy indices.
Shin et al. developed an algorithm to encrypt private medical images and transfer the encrypted images to improve the encryption capability and elapsed time. They used a fixed Mersenne prime number from the modified Rivest-Shamir-Adelman algorithm to test the encryption capability. Results showed the developed cryptographic algorithm provided improved encryption in medical ultrasound imaging compared to the comparable algorithm that without using the Mersenne prime number.
Pulmonary embolism and other pulmonary vascular diseases have been found associated with the changes in arterial morphology. Zhang et al. proposed a novel, fully automatic method that can extract pulmonary arterial tree from the computed tomographic pulmonary angiography images. The approach was based on fuzzy connectedness framework, combined with 3 D vessel enhancement and Harris corner detection. The new method achieved a mean accuracy of 92% when compared to manual reference, which was higher than 89% accuracy achieved by random forest method.
Percutaneous needle puncture operation is widely used in the image-guided interventions, including biopsy and ablation. MRI guidance has advantages of high-resolution soft tissue imaging and thermal monitoring during energy-based ablation. Zhao et al. designed a 5-DOF pneumatic needle puncture robot, with MRI-compatible cylinders, sensors and structure material. In addition, a hybrid fuzzy-proportional integral derivative (PID) controller was designed to adaptively adjust the PID parameters. Compared with the traditional PID control, the proposed hybrid fuzzy-PID control reported no overshoot and lower settle time/steady state error, indicating it can enhance the positioning accuracy and benefits the safety of percutaneous needle puncture operation.
Li et al. provided an assisted therapeutic system based on reinforcement learning (RL) for children with autism spectrum disorders. The integration of RL and convolutional neural network (CNN)-support vector regression was used to deal with the updating online of prediction model’s weights. The normalized emotion labels were evaluated by the therapists. Eleven children with autism were invited in the experiment and their facial video images were captured. Results demonstrated the feasibility of assisted therapeutic system in children with autism.
Zhu et al. aimed at creating a database of tactile information to provide guideline in defining minimally invasive surgery (MIS). They used the Fuji film Prescale pressure measuring system to measure the contact pressure and distribution at the jaws of laparoscopic grasping forceps. This data was then correlated with measured pressures at the forceps handles to understand the relationship between the surgeon’s actuating pressure and that on the organ being manipulated.
Chen et al. proposed a novel method for the complex scenarios classifications of EEG signals by combining fast Fourier transform (FFT) and support vector machine (SVM) methods, which improved the accuracy and comprehensive efficiency of multi-class EEG signal recognition, by employing data standardization preprocessing, feature extraction with FFT and principal component analysis (PCA), as well as the weighted k-nearest neighbor method. Experiments were conducted using public datasets of brainwave 0–9 digits classification. The precision rate, recall rate, accuracy and F1 score of the new developed method were 89%, 85%, 84% and 85%, respectively, which were better than the comparable methods.