AMD-Network: Automatic Macular Diagnoses of disease in OCT scan images through Neural Network Praveen Mittal, Charul Bhatnagar G.L.A University, Mathura, Uttar Pradesh, India Abstract Retinal optical coherence tomography scan images are used to diagnose retinal diseases using Convolutional Neural Network. One of the most beneficial of using Optical coherence tomography scan images is its non-invasive operation. Most ophthalmologists are using optical coherence tomography images for treating the retinal disorder in the human eye but due to its high cost of imaging, every patient can effort this imaging modality. Convolution Neural networks are now day giving lots of opportunities to classify various classes of images automatically. This method first removes noise from the images which get induced at the time of image capturing. Usually, Gaussian noise easily gets introduces in the transmission of images from one device to another. Speckle noise can be easily removed with the help of an average filter with a deviation of 0.7. Convolution Neural Network first trains the network with the help of activation functions like rectilinear Unit. The proposed method achieves 98.8% accuracy in the dataset of 50000 images Keywords Choroidal Neovascularization, Diabetic retinopathy, Diabetic Macular Edema, Glaucoma, Human Retinal Disease, Residual Network, VGG16. deviation of 0.7 and mask (9,9) could efficiently 1 Introduction remove noise from the scanned images [6]. The human eye plays a vital role in training the Ophthalmology is evolving by many research human brain. Retinal diseases such as diabetic scholars [7]. Eyes being a major part of the human macular edema, glaucoma, diabetic retinopathy, body require advanced techniques for detection and and Choroidal Neovascularization not only impact treatment of various diseases - OCT is one such the vision of the human eye but also made a great technique for imaging [8]. impact on learning the visual object's behavior [1]. Optical Coherence Tomography has become a Handling OCT images becomes a tedious task due useful imaging modality now a day due to its non- to various issues such as noise in the image, less invasive medical imaging method [2]. Retinal visibility, unnoticeable variation in intensity Optical Coherence Tomography allows between different layers of Retina, etc. OCT ophthalmologists to diagnose retinal diseases such images are used to examine various eye diseases as age-related macular degeneration and glaucoma and to know the condition of the retinal layout [9]. [3]. Therefore, it stings a very important role in the field of Ophthalmology [10]. Processing Optical coherence topographical images is not an easy task due to speckle noise present in Denoising is usually one of the major steps in the images at the time of scanning [4]. classification pre-processing [11]. There is many convolutional neural networks (CNN) based Many research scholars [5] gave techniques to algorithm as mentioned in [12] on a chest X-ray remove these speckle noise from optical coherence dataset to classify medical Images of pneumonia. tomography but Gaussians filter with a standard We have hooked upon OCT images for this process as they have a huge relevance when it comes to generating a schematic that defines multiple tissues ACI’2021: Workshop on Advances in Computational of the eye, called the Retina [13]. It is also used for Intelligence, February 25-27, 2021, Delhi, India EMAIL: praveen.mittal@gla.ac.in (P. Mittal); charul@gla.ac.in examination, by imaging the eyes of the patients (C. Bhatnagar) with various eye conditions such as Diabetic ORCID: 0000-0002-5331-7154 (P Mittal); Macular edema and diabetic retinopathy, etc [14]. Noise makes its way into the OCT images during © 2020 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). acquisition. The sensor and circuitry of CEUR Workshop Proceedings (CEUR-WS.org) a scanner or digital camera could also raise the problem of getting noise to the image [15]. Further film grain can also add to be a reason behind image images and work further on the results obtained noise. Never intended to be introduced in an image [19]. Here, Technique aim at diagnosing a various it degrades the quality of an image. Hence, disease that comprises the retina of the eye. The processing an image becomes necessary [16]. method needs to find the thickness of each layer Image processing is the procedure of improving the after finding different layers, to examine the eye for quality and information content of the original data. various diseases. Different layers may consist of Image enhancement, Restoration is amongst some different diseases which need to be diagnosed. relief approaches that are used to improve the What adds to the problem is that the gradient quality of an image. The latest research in [17] amongst the layers is decent and hence it further expresses the application of deep learning in becomes a tedious task to segregate the layers. In medical image processing. this paper, AMD-net is used that works on cross- validation which aims to figure the classification 2 Related Work hence, generating the desired classification or There are the number of classification work has clustering of images. been done for OCT images till date, but they are for classifying diseased eye from Normal eye. The following section describes the various works done in the classification of retinal Spectral Domain OCT (SDOCT) images. M. Treder et al. proposed a model in [5], [18], which is based on machine learning for retinal SDOCT image categorization for Age-related Macular Degeneration with the help of a dataset obtained from Heidelberg. The authors use the Inception-v3 model for Deep Convolutional Neural Network where starting layer got trained on ImageNet and the final layer got trained for taken dataset. Their work showed a good result for age- related macular degeneration diseases. Their work was only designed for age-related macular degeneration disease and Normal eye images Fig. 2. Accumulation of retinal fluid in layers [17] 3 Dataset Dataset of optical coherence tomographic retinal images which is used in this proposed work is freely available on the Kaggle repository. In the taken dataset there are four classes of images names Diabetic Macular edema, Diabetic Retinopathy, Glaucoma, and Healthy eye. Total 50000 images over which convolutional neural networks have been trained. 4 Proposed Work The proposed word showed the classification of retinal images with an accuracy of 98.8%. This work uses Neural Network to train the neurons in the network and provide a feature vector for the Fig. 1. OCT scanning of the human eye [18]. identification of features of the particular disease. The next step that comes is Segmentation. Segmentation is a process of dividing an image into regions. This technique is a mid-level processing technique. This further aims to segment the OCT Convolutn_two_dim(submatrics=16, Convolutn @5 (Convtn2Dm) (Null, 59, 59, 32) 4640 mask_shape=(5,5), S_rate=1, stuffing='valid', ___________________________________________________ act_funtn ='relunit', input_shape=size) Convolutn @6 (Convtn2Dm) (Null, 57, 57, 32) 9248 ___________________________________________________ Convolutn_two_dim (submatrics =16, mask_shape Maxpoollayer@3 (MaxPoolingtwodim (Null, 28, 28, 32) 0 ___________________________________________________ =(5,5), S_rate =1, stuffing ='valid', act_funtn Convolutn@7 (Convtn2Dm) (Null, 26, 26, 64) 18496 =''relunit ') ___________________________________________________ Convolutn@8 (Convtn2Dm) (Null, 24, 24, 64) 36928 MxPoolingLayer_two_dim(pool_sz=(4,4)) ___________________________________________________ Maxpoollayer@4 (MaxPooling2dm (Null, 12, 12, 64) 0 Convolutn_two_dim (submatrics =32, mask_shape ___________________________________________________ =(5,5), S_rate =1, stuffing ='valid', act_funtn loosingvalues@1 (Dropout) (Null, 12, 12, 64) 0 ___________________________________________________ =''relunit ') flattenofcurvature@1 (Flatten) (Null, 9216) 0 ___________________________________________________ Convolutn_two_dim submatrics =32, mask_shape compact_1 (Compact) (Null, 128) 1179776 =(5,5), S_rate =1, stuffing ='valid', act_funtn ___________________________________________________ compact_2 (Compact) (Null, 4) 516 ='relunit ') ============================================== Overall attribute: 2,153,756 MxPoolingLayer_two_dim (pool_sz=(4,4)) learnable attribute: 2,153,756 Non-learnable attribute: 0 Convolutn_two_dim (submatrics =64, mask_shape =(5,5), S_rate =1, stuffing ='valid', act_funtn =''relunit ') 5 Results and Analysis Convolutn_two_dim (submatrics =64, Implementation results are elaborated in Table 2 mask_shape =(5,5), S_rate =1, stuffing ='valid', for the hyper attribute of the experimented network. act_funtn ='relunit ') Table 2 Hyper attribute of the experimented network MxPoolingLayer_two_dim (pool_sz=(4,4)) Hyper Value Convolutn_two_dim (submatrics =128, attribute Damage ‘categorical mask_shape =(5,5), S_rate =1, stuffing ='valid', Tiny group size cross-entropy 200 act_funtn =''relunit ') Epoch 250 Convolutn_two_dim (submatrics =128, Premature 12 val_loss mask_shape =(5,5), S_rate =1, stuffing ='valid', Typical Barrier Was_opt_5 act_funtn='relunit') stuffing ‘valid’ Optimization ‘diffGrad’ MxPoolingLayer_two_dim (pool_sz=(4,4)) algorithm multiprocessing ‘Null’ Table 1. Classification of OCT Scan images as per the classes. ___________________________________________________ Layer (type) Output Shape Param # ============================================== Convolutn @1 (Convtn2Dm) (Null, 254, 254, 8) 80 ___________________________________________________ Convolutn @2 (Convtn2Dm) (Null, 252, 252, 8) 584 ___________________________________________________ Maxpoollayer@1 (MaxPoolingtwodim (Null, 126, 126, 8) 0 ___________________________________________________ Convolutn @3 (Convtn2Dm) (Null, 124, 124, 16) 1168 ___________________________________________________ Fig. 3 Training and testing values of the Network Convolutn @4 (Convtn2Dm) (Null, 122, 122, 16) 2320 on retinal images ___________________________________________________ Maxpoollayer@2 (MaxPoolingtwodim (Null, 61, 61, 16) 0 ___________________________________________________ Table 3 Confusion matrix for the proposed model Further, the classification time of the experimented convolutional neural network is 1.061 sec which is 19 sec less than the time taken by ResNet50 for the OCT Images set of retinal images. So if we talk about the total 23 time taken to process the retinal OCT images for DR 5 3 2 4 classification is 1.173 sec, which is less than the True label DME 3 254 2 4 time taken by ImageNet and ResNet50 on the same Glau- retinal OCT images. coma 21 2 237 6 Healt hy 1 1 1 219 7 References DR DME Glaucoma Healthy [1] Keel S, Lee PY, Scheetz J, et al. Feasibility and patient acceptability of a novel artificial intelligence-based screening model for diabetic retinopathy at endocrinology outpatient services: Precision = TP/(TP+ FP ) (1) a pilot study. Sci Rep 2018;8:4330. [2] Chen M, Wang J, Oguz I, et al. Automated Recall = TP/(TP+FN) (2) segmentation of the choroid in EDI-OCT images with retinal pathology using convolution neural F1-Score networks. Fetal Infant Ophthalmic Med Image = 2*(precision*recall)/( precision + recall) (3) Anal 2017;10554:177–84. [3] Worrall D, Wilson CM, Brostow GJ. Automated retinopathy of prematurity case Table 4 Precision-recall values for proposed model detection with convolutional neural networks, 2016. [4] Brown JM, Campbell JP, Beers A. Fully Precision recall f1-score support automated disease severity assessment and DR 0.798 0.998 0.980 250 treatment monitoring in retinopathy of prematurity DME 0.883 0.864 0.983 250 using deep learning. Proceedings Volume 10579, GLAUCOMA 0.890 0.780 0.931 250 Medical Imaging 2018: Imaging Informatics for Healthy 0.845 0.876 0.965 250 Healthcare, Research, and Applications, 2018. [5] S. A. Pai, N. Hussain, S. P. Hebri, A. M. accuracy 0.952 1000 Lootah, and M. A.Dekhain, "Volcano-like pattern macro avg 0.954 0.952 0.922 1000 in optical coherence tomography in chronic diabetic macular edema," Saudi Journal of Table 5 Learning time and validation time of Ophthalmology, vol. 28, pp. 157-159,41/2014. experimented model [6] J. Abelian, R. M. Baker, F. P. A. Coolen, R. J. Crossman, and A.R. Masegosa, "Classification Per epoch training time is 3600sec With binary classifications from a nonparametric predictive inference perspective," Type Time Computational Statistics & Data Analysis, vol. 71, training Time 355000 sec pp. 789-802, 31/2014. validation Time 125 sec https://doi.org/10.1016/j.compmedimag.2018.01.00 1 PMID: 29366655. [7] Sidibè D, Sankar S, LemaõÃtre G, Rastgoo M, Table 6 Processing time and classification time of Massich J, Cheung CY, et al. An anomaly proposed AMD-Net detection approach for the identification of DME patients using spectral-domain optical coherence Type Time tomography images. Computer Methods and Processing Time 0.112 sec Programs in Biomedicine. 2017; 139:109±117. Classification Time 1.061 sec https://doi.org/10.1016/j.cmpb.2016.11.001 PMID: Total Time 1.173 sec 28187882. [8] Hussain MA, Bhuiyan A, Ishikawa H, Smith RT, Schuman JS, Kotagiri R. An automated 6 Conclusion method for choroidal thickness measurement from Enhanced Depth Imaging Optical Coherence The table shown above depicts the results in the Tomography images. Computerized Medical processing time of the experimented convolutional Imaging and Graphics. 2018; 63:41±51. neural network as 0.112 sec which is less than the [9] L. Ngo, G. Yih, S. Ji, and J. H. Han: A study processing time of the previous research till now. on automated segmentation of retinal layers in optical coherence tomography images.4th Deep Learning Methods”, Multimedia Tools and International Winter Conference on Brain- Applications (MTAP), 80(1), pp. 1-32. Computer Interface (BCI), 1–2 (2016) DOI: https://doi.org/10.1007/s11042-020-09818-1. 10.1109/iww-bci.2016.7457465. [15] Gurpreet Kaur, Prateek Agrawal, [10] J. Oliveira, S. Pereira, L. Gonçalves, “Optimisation of Image Fusion using Feature M.Ferreira, and C. A. Silva: Sparse high order Matching Based on SIFT and RANSAC”, Indian potentials for extending multi-surface segmentation Journal of Science and Technology, 9(47), pp 1-7, of OCT images with drusen.37th Annual 2016. International Conference of the IEEE Engineering [16] Abhishek Sharma, Prateek Agrawal, Vishu in Medicine and Biology Society (EMBC), 2952– Madaan and Shubham Goyal, “Prediction on 2955 (2015)DOI: 10.1109/embc.2015.7319011. Diabetes Patient’s Hospital Readmission Rates”, [11] N. Deepa, SP.Chokkalingam, "Deep 3rd International Conference on Advances Convolutional Neural Networks (CNN) for Informatics on Computing Research (ICAICR’19), Medical Image Analysis", International Journal of pp. 1-5, Jul 2019, ACM-ICPS. Engineering and Advanced Technology (IJEAT), [17] P Mittal, C Bhatnagar, Automatic ISSN: 2249–8958, Volume-8, Issue-3S, February classification of retinal pathology in optical 2019. coherence tomography scan images using [12] Samir S. Yadav and Shivajirao M. Jadhav, convolutional neural network, Journal of Advanced "Deep convolutional neural network-based medical Research in Dynamical and Control Systems 12 image classification for disease (3), 936–942. diagnosis",https://doi.org/10.1186/s40537-019- [18] P Mittal, C Bhatnagar, Detecting outer edges 0276-2. in retinal OCT images of diseased eyes using graph [13] Vishu Madaan, Aditya Roy, Charu Gupta, cut Method with weighted edges, Journal of Christian Bologna, Radu Prodan, "XCOVNet: Advanced Research in Dynamical and Control Chest X-ray Image Classification for COVID-19 Systems 12 (3), 943–950 Early Diagnosis using Convolution Neural [19] P.Mittal, Automatic segmentation of Networks", New Generation Computing, Springer, pathological retinal layer using eikonal equation, 2021. 11th International Conference on Advances in [14] Prateek Agrawal, Deepak Chaudhary, Vishu Computing, Control, and Telecommunication Madaan, Anatoliy Zabrovskiy, Radu Prodan, Dragi Technologies, ACT 2020, 2020, pp. 43–49 Kimovski, Christian Timmerer, “Automated Bank Cheque Verification Using Image Processing and