Study and Analysis of Fluid Filled Abnormalities in Retina Using OCT Images

Abstract

Visual impairment is one of the most regularly happening infections in human. The reason being variation from the normal in the different layers of retina because of strange measure of liquid either abundance aggregation or shortage. This paper targets recognizing and assessing the different abnormalities that could be earlier stages to visual deficiency. The proposed target is achieved by means of implementation using Digital Image Processing Technique, starting from preprocessing to classification at various stages. Not restricting to binary classification as normal or abnormal, the proposed system also extends its capacity to classify the input image as Cystoid Macular Edema (CME), Choroidal Neo Vascular Membrane (CNVM), Macular Hole (MH) and normal images. The preprocessing methodology implemented filters to remove the speckle noises which are most common in ultrasound-based imaging system. Random forest classifier was utilized for classifying the input features and also seems to be promising on par with the various existing methodologies.

1. Introduction

Biomedical Imaging and automations in medicine has become an emerging field due to the need of precision, accuracy and storage/retrieval capability of the data being handled. Automations in the field of ophthalmology has also proven to be an emerging need of the hour as more accurate diagnosis of small changes in the retinal layers shall prevent the vision impairment. Such variations are commonly initiated owing the alterations in fluid pattern of the retinal coverings. The fluid may become excess or deficit based on the specified abnormality and due to this the patient may suffer a loss of vision. In order to identify this fluid pattern more accurately ultrasound-based imaging modality, namely, developing imaging techniques such as Optical Coherence Tomography (OCT) Ghastly Domain Optical Coherence Tomography (SD-OCT) mentioned in [1], which unmistakably differentiates numerous infections in different layers of the retina [2]. An inner layer of an eye is Retina which changes over the occurrence bright flag into neural sign, which are conveyed to the mind. It comprises of different shades, poles and cones are in charge of blurred light and shading dreams individually. Retina has a few layers of neurotic and physiological significance. Some harms in the retina layers lead to few more hazard variations from the norm including vision loss. Prior examination of OCT Images focused on separation of Intraretinal layers [3, 4, 5, 6, 7, 8, 9, 10, 11, 12], division of fluid engaged layers, [13, 14] and optical circle is a fundamental one.

Medical Image Processing classification assumes a remarkable work towards familiar proof and outcome of variations are observed from the model in different imaging modalities. OCT employs ultrasound waves as a source and transceives the corresponding pictures of different layers of retina [15]. OCT is utilized in 3D reconstructible picture of retina which is required for better comprehension of real investigation, liquid-based variation from the norm. Priori stages to visual deficiency incorporate modifications in the liquid example of the different retinal layers. Scarcely any such anomalies are cystoids macular edema (CME), wherein the macular layer gets gathered with abundance of liquids, Choroidal Neo Vascular Membrane (CNVM), a fresh recruits’ vessels are framed and liquid substance winds up more in the Choroidal layer, and Macular Hole (MH), wherein the macular layer progresses towards becoming shortfall of liquids. Various methodologies and implementation of the proposed system is explained in the below section.

2. Methods and materials

The overall flow diagram of the entire methodology could be clearly understood from Figure 1.

Figure 1.

Proposed methodology.

Images are acquired using SD OCT Device in digital format and provided as input to our system. Image inputs are exposed to preprocessing which includes grayscale conversion and noise removal. OCT practices ultrasound as a wellspring of imaging, the pictures are increasingly inclined to spot commotions, in nature these are multiplicative [16, 17, 18, 19, 20]. So as to dispense with/limit these commotions, homomorphic channels were utilized to kill the spot clamors. In spite of the fact that different channels are available. In the wake of preprocessing, so as to seclude the retinal layers, various methodology of division [19, 20, 21, 22] is listed out. This assurance the retinal layers of intrigue remain disengaged from the repetitive foundation for additional handling of the image. This dynamic division calculation, the Gradient Vector stream calculation of division in the framework of projected and assessed. The division calculation figures the dissemination of slope vectors of dim level or paired edge map which is gotten from the info preprocessed picture. Therefore, the yield is normal give an unmistakable perspective on the obligatory retinal layers with liquid founded irregularities as opposed to considering the whole picture information for further preparing.

Different highlights of the divided picture were extricated among which they chose highlights are Skewness, Entropy, Variance and Energy which are the factual highlights which are utilized as info information in characterization of the pictures as typical and strange. Critical contrasts could be found in the highlights of ordinary and unusual pictures and consequently these are utilized in further arrangement. (A variance image, an image of variances that is the squares of the standard deviations, of input or output images). If a variance set is delivered, all pixel values of the variance image must be the pixels of inset which may have any value.

The separated highlights are part for preparing and testing/approval of the classifiers in a proportion of 3:1 for accomplishing a superior generally speaking execution of the classifiers. The highlights are exhibited to Random Forest Classifier so as to get the yield in paired structure either as Normal or Abnormal. Random Forest Classifier assesses the information include into various classes and plays out a democratic check to decide the most extreme likelihood in having a place of the information highlight vector. The classifier checks for the lion’s share casting a ballot under each woodland and characterizes the info occurrence dependent on the most extreme votes. Random forest includes of a huge number of discrete choice trees that work as a congregation.

Endless supply of the component information, above all clarified classifiers, the yields of the equivalent are arranged to recognize untrue positive, True negative, genuine positive and untrue negative [23, 24]. From these qualities’ affectability, explicitness and thus the general framework precision/execution is being assessed for every one of the classifiers independently and this outcome is utilized to recognize the best classifier among the proposed calculations.

3. Result

The new proposed procedure was realized for 114 Images which were obtained from Topcon and Zeiss OCT gadgets. Among the accessible dataset of 27 Normal Images and 87 Abnormal Images, in absolute 70 level of the pictures were used for preparing and 30 rate where utilized for testing and validations. Right off the bat the picture changes over gray scale and exposed to preprocessing utilizing various filters for evacuation of speckles. While the original input image and gray level image could be seen in Figures 2 and 3 respectively, the outputs of various filters evaluated is shown below in Figure 4.

Figure 2.

Input image.

Figure 3.

Grayscale image.

Figure 4.

Output of various speckle reduction filters. (a) Mean filter 3*3, (b) mean filter 5*5, (c) mean filter 7*7, (d) frost filter, (e) adaptive smoothing filters, (f) Gaussian filters 3*3, (g) Gaussian filters 5*5, (h) Gaussian filters 7*7, (i) bilateral filters, (j) anisotropic diffusion filters, (k) homomorphic wiener filters 3*3, (l) homomorphic wiener filters 5*5.

When most of the filters just average the data instead of removing the speckles, the computational results show that Homomorphic wiener filter is comparatively better, as the filter is converting the multiplicative noises into additive and hence noise removal is much efficient.

Preprocessed images are exposed to segmentation utilizing Gradient Vector Flow calculation. From the segmented image, different highlights like, Skewness, Entropy, Energy, and Variance were extricated and these highlights spoke to have a critical striking distinction among the ordinary and anomalous pictures. With these highlights as features to the classifiers, the data is being presented to the random forest network classifier whose various statistical values are being tabulated in Tables 1 and 2. The various parameters of statistics include True Positives, True Negatives, Untrue Positives and Untrue Negatives, Specificity, Accuracy, and Sensitivity.

Condition	TP (%)	FP (%)	FN (%)	TN (%)
Abnormal	87.50	12.50	96.15	3.85
Normal	96.15	3.85	87.50	12.50

Table 1.

Performance output of random Forest classifier.

Classifier	Sensitivity	Specificity	Accuracy
Random Forest	87.50	96.15	91.82

Table 2.

Performance analysis of overall system.

From the general outcomes it is reasonable that Random Forest classifier is by all accounts equally more hopeful than different classifiers. These results seem to be similarly efficient and comparable with the previous related works.

4. Discussion

The Random Forest classifier seems to be comparatively efficient which is shown in the obtained results, when compared to the other classifiers, comparing to the earlier related works [25, 26, 27], the proposed system seems to be equally efficient towards classifying the input image as abnormal and normal. The overall system parameters like specificity, sensitivity, and accuracy seems to be more hopeful when compared with the existing methodologies as cited. By expanding the range of abnormalities and also identifying age related disorders can be further enhanced.

5. Conclusion

The proposed framework, and calculations of arrangement, is by all accounts proficient. The degree of precision accomplished demonstrates that the framework can likewise be used for useful ramifications. The effectiveness of the proposed classifier frameworks should be assessed for explicit variation from the norm-based arrangement like Cystoid Macular Edema, Choroidal Neo Vascular Membrane, Epi Retinal Membrane, Macular Hole, Age related Macular Degeneration in future, in view of the real execution could be precisely examined in detail. On definite issue-based order study, the created framework can likewise be coordinated with the OCT gadgets progressively to consequently remark on the pictures that are given as the contribution to the framework.

The developed system was implemented and evaluated with MATLAB and the validation of the system was done with reference to Ophthalmologists. The comments of the ophthalmologists on the proposed system is attached. Though the overall accuracy of the system matches the earlier works, significant improvisation could be seen with the number of abnormalities that had been included in the process of classification. With overall system accuracy of around 91.65% the proposed system is significantly in match with the existing systems. The accuracy, sensitivity, specificity and Youden’s index are acceptable for the system proposed and has been cross validated for its performance by Ophthalmologists and found to be satisfactory. The obtained results in terms of accuracy of 91.65% are closer with the related works whose accuracies are closer to the values obtained by the proposed system Reza Rasti et al. [16]. The obtained performance metrics shows that the acquired accuracy is remarkably higher than the works of Philipp Seebock et al. of 81.4%, which focused to limited abnormalities. The comparison could be understood from the representation in Table 3.

Metrics	Proposed	Philipp Seebock et al.
Accuracy	91.65%	81.4%
Number of disorders	3	1

Table 3.

Comparison of performance.

Earlier works focused on morphological parameters alone which restricted the number of disorders taken into consideration for categorization. Apart from the morphological parameters of bulkiness, compactness and convexity, the evaluated system also used Zernike moments which is widely varying feature for the disorders chosen. The overall number of features considered has been comparatively reduced with significant accuracy, thereby optimizing the features used for the process of classification. Based on the excess or deficit of fluids in the retinal layers, the overall features also varied accordingly, so as to enable the classifier to remain comparatively more accurate for the proposed system. Zernike moments are most commonly used shape descriptors and were hence used for the proposed application, in order to detect the shape-based changes that occur in retina due to accumulation or lack of fluid within the retinal layers. The classifier that has been used is a basic supervised classifier, yet more appropriate for the application proposed. The average Youden ‘s Index shows that the algorithm proposed is reliable in retinal analysis and could be used in automated analysis of OCT Image analysis. The proposed system could be further extended for other disorders in retina and integrated with OCT Device as an additional software for instantaneous evaluation of the retinal disorders and the therapeutic efficiencies. The developed system is promising for the selected application and has been evaluated with comparatively higher number of input samples, which restricted its evaluation until the process of segmentation and did not focus on classification further. The overall performance indices are also satisfactory and matchable with the existing results derived from similar works as evaluated by the Ophthalmologist. The developed system has also covered a significant number of fluid related disorders which are caused due to excessive or deficit fluids within the layers of retina. As OCT is an efficient tool for detection of prior stages of blindness, the proposed algorithm remains an expert system for earlier identification and accurate evaluation of the retinal fluid volumes.