Features extraction using encoded local binary pattern for detection and grading diabetic retinopathy

Introduction

Reliable computer diagnosis of diabetic retinopathy (DR) is needed to rescue many with diabetes who may be under threat of blindness. This research aims to detect the presence of diabetic retinopathy in fundus images and grade the disease severity without lesion segmentation.

Methods

To ensure that the fundus images are in a standard state of brightness, a series of preprocessing steps have been applied to the green channel image using histogram matching and a median filter. Then, contrast-limited adaptive histogram equalisation is performed, followed by the unsharp filter. The preprocessed image is divided into small blocks, and then each block is processed to extract uniform local binary patterns (LBPs) features. The extracted features are encoded, and the feature size is reduced to 3.5 percent of its original size. Classifiers like Support Vector Machine (SVM) and a proposed CNN model were used to classify retinal fundus images. The classification is abnormal or normal and to grade the severity of DR.

Results

Our feature extraction method was tested on a binary classifier and resulted in an accuracy of 98.37% and 98.84% on the Messidor2 and EyePACS databases, respectively. The proposed system could grade DR severity into three grades (0: no DR, 1: mild DR, and 5: moderate, severe NPDR, and PDR). It obtains an F1-score of 0.9617 and an accuracy of 95.37% on the EyePACS database, and an F1-score of 0.9860 and an accuracy of 97.57% on the Messidor2 database. The resultant values are dependent on the selection of (neighbours, radius) pairs during the extraction of LBP features.

Conclusions

This study’s results proved that the preprocessing steps are significant and had a great effect on highlighting image features. The novel method of stacking and encoding the LBP values in the feature vector greatly affects results when using SVM or CNN for classification. The proposed system outperforms the state of the artwork. The proposed CNN model performs better than SVM.

Histogram equalization using a selective filter

Many popular modern image processing software packages implement a naïve form of histogram equalization. This implementation is known to produce histograms that are not truly uniform. While exact histogram equalization techniques exist, these may produce undesirable artifacts in some scenarios. In this paper we consider the link between the established continuous theory for global histogram equalization and its discrete implementation, and we formulate a novel histogram equalization technique that builds upon and considerably improves the naïve approach. We show that we can linearly interpolate the cumulative distribution of a low-bit image by approximately dequantizing its intensities using a selective box filter. This helps to distribute the intensities more evenly. The proposed algorithm is subsequently evaluated and compared with existing works in the literature. We find that the method is capable of producing an equalized histogram that has a high entropy, while distances between similar intensities are preserved. The described approach has implications on several related image processing problems, e.g., edge detection.

Deep learning, reusable and problem-based architectures for detection of consolidation on chest X-ray images

BACKGROUND AND OBJECTIVE

In most patients presenting with respiratory symptoms, the findings of chest radiography play a key role in the diagnosis, management, and follow-up of the disease. Consolidation is a common term in radiology, which indicates focally increased lung density. When the alveolar structures become filled with pus, fluid, blood cells or protein subsequent to a pulmonary pathological process, it may result in different types of lung opacity in chest radiograph. This study aims at detecting consolidations in chest x-ray radiographs, with a certain precision, using artificial intelligence and especially Deep Convolutional Neural Networks to assist radiologist for better diagnosis.

METHODS

Medical image datasets usually are relatively small to be used for training a Deep Convolutional Neural Network (DCNN), so transfer learning technique with well-known DCNNs pre-trained with ImageNet dataset are used to improve the accuracy of the models. ImageNet feature space is different from medical images and in the other side, the well-known DCNNs are designed to achieve the best performance on ImageNet. Therefore, they cannot show their best performance on medical images. To overcome this problem, we designed a problem-based architecture which preserves the information of images for detecting consolidation in Pediatric Chest X-ray dataset. We proposed a three-step pre-processing approach to enhance generalization of the models. To demonstrate the correctness of numerical results, an occlusion test is applied to visualize outputs of the model and localize the detected appropriate area. A different dataset as an extra validation is used in order to investigate the generalization of the proposed model.

RESULTS

The best accuracy to detect consolidation is 94.67% obtained by our problem based architecture for the understudy dataset which outperforms the previous works and the other architectures.

CONCLUSIONS

The designed models can be employed as computer aided diagnosis tools in real practice. We critically discussed the datasets and the previous works based on them and show that without some considerations the results of them may be misleading. We believe, the output of AI should be only interpreted as focal consolidation. The clinical significance of the finding can not be interpreted without integration of clinical data.

Assessing the impact of choosing different deformable registration algorithms on cone-beam CT enhancement by histogram matching

Background

The aim of this work is to assess the impact of using different deformable registration (DR) algorithms on the quality of cone-beam CT (CBCT) correction with histogram matching (HM).

Methods and materials

Data sets containing planning CT (pCT) and CBCT images for ten patients with prostate cancer were used. Each pCT image was registered to its corresponding CBCT image using one rigid registration algorithm with mutual information similarity metric (RR-MI) and three DR algorithms with normalized correlation coefficient, mutual information and normalized mutual information (DR-NCC, DR-MI and DR-NMI, respectively). Then, the HM was performed between deformed pCT and CBCT in order to correct the distribution of the Hounsfield Units (HU) in CBCT images.

Results

The visual assessment showed that the absolute difference between corrected CBCT and deformed pCT was reduced after correction with HM except for soft tissue-air and soft-tissue-bone interfaces due to the improper registration. Furthermore, volumes comparison in terms of average HU error showed that using DR-NCC algorithm with HM yielded the lowest error values of about 55.95 ± 10.43 HU compared to DR-MI and DR-NMI for which the errors were 58.60 ± 10.35 and 56.58 ± 10.51 HU, respectively. Tissue class’s comparison by the mean absolute error (MAE) plots confirmed the performance of DR-NCC algorithm to produce corrected CBCT images with lowest values of MAE even in regions where the misalignment is more pronounced. It was also found that the used method had successfully improved the spatial uniformity in the CBCT images by reducing the root mean squared difference (RMSD) between the pCT and CBCT in fat and muscle from 57 and 25 HU to 8HU, respectively.

Conclusion

The choice of an accurate DR algorithm before performing the HM leads to an accurate correction of CBCT images. The results suggest that applying DR process based on NCC similarity metric reduces significantly the uncertainties in CBCT images and generates images in good agreement with pCT.

Discrimination of spatial phase: The roles of luminance distribution and attention

We can easily discriminate certain phase relations in spatial patterns but not others. Phase perception has been found different in the fovea vs. periphery, and for single patterns vs. textures. Different numbers of mechanisms have been proposed to account for the regularities of phase perception. In this study, I attempt to better understand the mechanisms behind discrimination of spatial phase. In order to reveal the role of luminance cues, I use histogram matching of patterns with different phases. Possible effects of attention were studied using visual search experiments with varied stimulus set size. Simple and compound Gabor patches, broadband lines and edges, and textures composed of those patterns were used as stimuli. The experiments indicate that phase discrimination is mediated by two mechanisms. The first uses luminance differences and operates pre-attentively, in parallel across the visual field. The second compares relative positions of dark and bright segments within an image, and is strictly limited by capacity of attention.

Automatic dynamic range adjustment for ultrasound B-mode imaging

In medical ultrasound imaging, dynamic range (DR) is defined as the difference between the maximum and minimum values of the displayed signal to display and it is one of the most essential parameters that determine its image quality. Typically, DR is given with a fixed value and adjusted manually by operators, which leads to low clinical productivity and high user dependency. Furthermore, in 3D ultrasound imaging, DR values are unable to be adjusted during 3D data acquisition. A histogram matching method, which equalizes the histogram of an input image based on that from a reference image, can be applied to determine the DR value. However, it could be lead to an over contrasted image. In this paper, a new Automatic Dynamic Range Adjustment (ADRA) method is presented that adaptively adjusts the DR value by manipulating input images similar to a reference image. The proposed ADRA method uses the distance ratio between the log average and each extreme value of a reference image. To evaluate the performance of the ADRA method, the similarity between the reference and input images was measured by computing a correlation coefficient (CC). In in vivo experiments, the CC values were increased by applying the ADRA method from 0.6872 to 0.9870 and from 0.9274 to 0.9939 for kidney and liver data, respectively, compared to the fixed DR case. In addition, the proposed ADRA method showed to outperform the histogram matching method with in vivo liver and kidney data. When using 3D abdominal data with 70 frames, while the CC value from the ADRA method is slightly increased (i.e., 0.6%), the proposed method showed improved image quality in the c-plane compared to its fixed counterpart, which suffered from a shadow artifact. These results indicate that the proposed method can enhance image quality in 2D and 3D ultrasound B-mode imaging by improving the similarity between the reference and input images while eliminating unnecessary manual interaction by the user.