View of Identification of Proper Machine Learning Classification Model based on Image Annotation Technique

(1)

Annals of R.S.C.B., ISSN:1583-6258, Vol. 25, Issue 4, 2021, Pages. 11135 - 11156 Received 05 March 2021; Accepted 01 April 2021.

11135 http://annalsofrscb.ro

Identification of Proper Machine Learning Classification Model based on Image Annotation Technique

Ms. Kavitha S. ^#1, Dr. Vidyaathulasiraman^*2

# Research scholar, Department of Computer Science,Periyar University, Salem, India.

*Assistant Professor and Head, Department of Computer Science,Government Arts and Science College for Women, Bargur,India

1[email protected], ²[email protected]

ABSTRACT

The ovary is a complex organ, the detection of the ovarian cancer is done at an early stage so that the death rate of women may decrease to certain level. The woman of old age have chance for suffering the severe illness of ovarian cancer. Depends upon the research 7^th mortality rate for woman is this ovarian cancer and this one is the 5^th normal cancer across the globe. By the use of ANN Artificial neural networks, lot of innovators classified the ovarian cancer. Decision making among the doctors believed in accuracy classification as the efficient factor. Accurate and Early diagnosis decreases the rate of mortality and secures life. In this paper a new annotated ovarian image classification using FR-CNN(fast region-based CNN) is proposed on segmented ROI basis. Input images are classified into three kinds they are epithelial, germ and stroma cells. Preprocessing and segmenting the images and then the process of annotation is proceded by FR- CNN. This work contrasts the features of annotation process and features which are trained in FRCNN manually for the purpose of classification which is region based. This will guide in the process of examining the increase accuracy. Completing the FRCNN training in region- based through the combination of classifiers like SVC- Support vector and Gaussian Naives Bayes. Because of increased indexing of data, the ensembling method was utilized in feature classification. Results of simulation provides the accurate part of input image for detecting ovarian cancer.

Keywords: Ovarian cancer, annotated image classification, FR-CNN (Fast Region-based CNN), ROI (Region

of Interest), SVM, Gaussian NB,Accuracy.

1. INTRODUCTION

In world, ovarian cancer is 2^nd leading cancer which affects about 2% of female over their lifetime. If it is diagnosed in the earlier stage, it has 90% survival rate. Many research reports after investigation presents that early symptoms and indicationsof ovarian cancer are not clear [1]. For ovarian cancer, medical experts face several problems in producing cancer-screening guide-lines, there exists no single known cause or mark which leads to make it as silent killer. Research reports show that 90% of patients have symptoms long back before diagnosed [2]. Further, many patients experience numerous tumor metastasis, treatment cycles and disease recurrences. In female genital tract, while considering endometrial and cervical cancer, ovarian cancer

(2)

is 3^rdcommon cancer in Taiwan [3]. Figure 1 represents cervical, endometrial and ovarian cancer incidence ratesin Taiwan from 1979 to 2013.

According to recent case studies related to control, women have many persistent or frequent non- specific symptoms. Symptoms are fullness or appetiteloss, improved frequency or urinary urgency, abdominal or pelvic pain, weight loss, changes in bowel habit or fatigue. Based on the examination of general databases, women between 45 to 70 years oldhave a consultation with their general practitioner each year with any of the above symptoms, which means that these are common symptoms. For doctors, this gives diagnostic issue, low incidence ovarian cancer and in general every, 3-5 years, there is one woman with ovarian cancer by UK general practitioner and for diagnosis, low „+ve‟ predictive values and lack of clear procedures. From the survey investigation of UK patients, 36% of women ovarian cancer diagnosis presents to physician with symptoms before diagnosis. To improve diagnostic accuracy, reduce human resources and costs, in medical analysis machine learning was introduced. Machine learning (ML) is the field of Artificial Intelligence (AI) which is developed by using computer programs based on tasks and performance measures [5].

Significant characteristic of intelligent behavior is learning ability. Machine learning is a study of computational method to enhance performance on some tasks. Its aim may be technical, theoretical or cognitive. In technical analysis, for knowledge based systems knowledge acquisition process is automated. In theoretical analysis, learning method characteristics like limitations and scope are considered. It is inherently interdisciplinary area similar to AI. In the field of ML, statistic is widely used. Based on different methods such as knowledge representation, learning strategy and domain application, ML is classified. It consists for 4 major paradigms. They are neural networks, genetic algorithm and analytic, inductive and instance based learning. These all methods having same goal to enhance performance of few tasks which is done by determining and exploiting regularities in training data.

The issues are formulated into suitable ML methods by selecting, determining, gathering training data, estimating and fielding the learned knowledge.For learning process, training data forms basis. Learning technique only determines concepts in training data. In problem domain, selected instances and attributes have various situations. For applying ML method, training data construction is significant which has involvement of both domain expert and knowledge engineer. To describe training cases is significant task in determining learning data representation in which attributes are used. Hence larger attribute has more data than smaller one which is not required to produce good results. For quantitative ones by assigning consecutive numerical scores to ordinal category and nominal transferred to binary dummy attributes, ordinal attributes are used.

Quantitative attributes are coded when learning needs discrete feature base. From large real-world dataset or set of instances from expert, training data is „raw‟ sample which is extracted. Larger samples statistically provide reliable picture of task to be learned, but it is not optimum for the purpose of learning. Use of performance measure is objective way for calculating learning output. To characterize classifier performance, most measure called accuracy is used [6].

(3)

For further classification, training stage of ovarian cancer detection from same objects/patterns/class is grouped together and used as reference. In training stage, same features were extracted and compared with obtained references. Classification performance depends on the extracted features. This paper aims to extract salient features and classify data whose results are used to develop a novel method cancer classification.

On knowledge basis, the symptoms of patients are classified to some particular group of diseaseby physicians. In this study, for ovarian disorders, learning classification model is learning task. By analyzing data, general classification method was learned. Using attribute vectors (features or variables), training data containing cases like objects or instances are described. It may be quantitative or qualitative. Mutually exclusive cases and class information are used for learning in supervised learning. If all cases with identical attribute vector from same class, attributes are suitable for classification task.

2. LITERATURE SURVEY

Ozkanİnika and Ayse Ceyhan presented automatic ovarian tissue follicles counting method based on CNN. This method attains accuracy of 86.68% of primordial, 95.34% primary, 97.05% preantral, 97.67%

secondary and 100% tertiary relating to expert counting. This method has accuracy of 95.35%. To analyze histopathological and histological images, this method was used. In cancer diagnoses of various tissues, it gives high accuracy. Automatically, image analysis was performed, manually image capture with microscope is performed [8].

Asha.L.Wankhededetected follicles growth, different kinds of ovarian follicle cell of fish images. This paper focused on SURF detection method which provided better detection result. It was the best method which was based on Haar wavelet response and with integral images it was evaluated significantly [9].

Work mentioned in [10], is follicle cell-oocyte interaction and from a theoretical point, one of every 2 structural differentiations of follicle cell/oocyte interface gap junctions and follicle cell microvilli potentially triggeredvitellogenesisinception. Gap junctions permitted regulatory molecule passage, transferred from follicle cells to oocytemanaged coated pits assembly on oocyte plasma membrane.

In [11], the authors focused in patterning follicle cell epithelium in anteriorposterior axis during Drosophila oogenesis. Methods used for this patterning was Fly stocks, Staining procedures, Generation of clones, Cell lineage analysis and results were the terminal follicle cell populations are similar prior to gurken signaling. During oocytes of Xenopus laevisdevelopment which establishes junctional contact with enveloping follicle cells proposed in [12]. Alternatively, these junctions have defined as gap junction and desmosomes. To experiment follicle cells detection using Digital Image Processing method, by using this literature surveys of this work.

(4)

Miao Wu, Chuanbo Yan presents Deep Convolutional Neural Networks (DCNN) based on AlexNetfor classifying various kinds of ovarian cancers automatically from cytological images. This structure comprises of 5 convolutional layers, 3 max pooling layers and 2 fully connected layers. Then classify trained model into two categories. One is original and other is augmented data image which has image rotation and enhancement. By using method of 10-fold cross validation, testing results are produced in which its accuracy is enhanced nearly by 6% with the use of augmented images as training data. For classifying ovarian cancers from cytological images, this method was very useful [13].

Uroosa Shafi and Sugandha Sharma presented identification and detection of ovarian cancer by image selection. Through various stages by acquiring images, this research helped to detect images. By various operators with MRI images, optimization method is used for extraction and selection of features. Using multi- layer CNN, detection and classification of ovarian cancer was performed. Using selection and extraction of features, classification process was done. This method was used to emphasis on comparing and calculating performance metrics. Hence image quality performance was enhanced with the parameters peak signal to noise ratio. Based on extracted features, low error rate, high accuracy and low processing time wereobtained [14].

Vasavi G and Jyothi S concentrated on ovarian growth, fibroids, poly sore ovary ailment and various problems found with gynaecoid problems. For patients, correct choice was taken on time, ovarian sores were legitimately described. Based on tumor width, treatment of choices was viewed. By using instrument ultrasound estimate, tumor distance was investigated and calculated. To apply best treatment on patients, evaluation of multi-dimensional parameter examination was required. On suspected organs, to identify precise heterogeneous structures and harmful development fake insight was used [15].

Schorge J. O and Modesitt S. C.carried an investigation on surveillance and screening of ovarian cancer. For avoiding ovarian malicious growth in high chance women, riskminimizing salpingo- oophorectomy was the significant method. Further, critical threat was reduced with inclusive community where oral contraceptives were utilized. Before determining the early illness, up to 89% patients have side effects. If more experience to restore was acquired by physicians, it may encourage to treat patients. If gigantic exertion is not known, no proof exists for normal screening of ovarian cancer which was either higher risk or every community with sonograms, serum markers, or pelvic investigations reduces mortality. In early identification, if any boards of markers or novel biomarkers have clinical utility needs further assessment to decide [16].

P.S. Hiremath and Jyothi R. Tegnoordescribed ovarian classification method for ultrasound images.

For preprocessing stage, this method uses contourlet transform and for follicle segmentation, active contour without edges was employed. Follicle was detected with fuzzy inference rules. After detection, ovary was divided into 2 parameters such number N and size S of follicles. To diagnose whether ovary is cystic, normal

(5)

or polycystic, fuzzy classifier was utilized. Experimental results provided better agreement with ovarian classes manual detected by experts to determinethe efficiency of this method. This method was used for ovarian abnormality classification which provided evidence in medico-legal cases based on infertility treatment as well as drug susceptibility tests while manufacturing the drugs [17].

Devesh D. Nawgaje, Rajendra D. Kanphade industrialized ovarian cancer image segmentation approach to select threshold, genetic method is used. Based on variance between background and target, GA adjusts mutation probability and crossover probability automatically in computational process. For diagnosis purpose, experimental data produced segmentation accuracy which helped doctors. Further, hardware implementation of this method on DSP TMS320C6713 was achieved successfully [18].

3. RESEARCH METHODOLOGY

The development in medical images provides more useful information hidden in image pixels where the medical practice/radiologists face difficulties to identify the power of diseases. Several approaches are being developed by the emerging researchers to solve this issue. The clinicians and radiologists show their interest in employing machine learning approaches particularly in detecting cancer. CAD methods are non- invasive butcost effective. Moreover, algorithms addressing heterogeneous medical image data are in demand.

Hence, this research work attempts to employ machine learning approaches in detecting ovarian cancer.

Thearchitecture of this research methodology is illustrated in figure 1.

Figure.1 Architecture for proposed system

Medical image dataset

(JPEG format)

Input image

preprocess

Testing Object detection

Training model

1. Resizing the image 2. Noise removal Preprocessing

Manual segmentation prescribed by physician

Faster- RCNN

Performanc e measures

Classification Gaussian SVC Annotation

Segmentation

Input img.

Epithelial OVC

Germ cell OVC Stroma OVC

Resize d image

Segmented image

(6)

The input image was classified into 3 kinds they are stroma, germand epithelialcells.Initially the images are preprocessed for noise removal and filtering. This preprocessed image has been manually annotated and trained using normal training model.For compensation of manual annotation here this research is done using the neural network known as Fast Region Convolution Neural Network (FR-CNN).From the trained image and manually segmented image, object is detectionusing FR-CNN. Here both features have been annotated based on region since the convolution is done for detecting edge.Contextual features have been annotated by image segmentation. The ratio of disease detection is lower manually and detection using computer aided diagnosis is higher in accuracy. Once FR-CNN is applied, Gaussian NB (Naives Bayes) and SVC (Support vector classifier) are used or classification.

3.1 Object Detection using FR-CNN

Fast Region CNN (FR-CNN) uses region proposal algorithm such as selective search to propose an estimated location of objects in an image. It is intuitive that the features extracted by CNN‟s are finally used to classify and give bounding boxes of the images. Thus, these extracted features have the information required to detect the objects in the image. FR-CNN architecture is built upon this observation. The region proposal algorithm (RPA) is replaced with region proposal network (RPN) that gives an estimate of regions with objects.

This RPN is based on CNN and gives region proposal from the extracted features of CNN.

FR-CNN architecture thus is a combination of RPN which proposes regions and uses these proposed regions to give final bounding boxes of the object. Since both FR-CNN and RPN requires a CNN based feature extractor to perform almost similar task (in the end task of RPN is to give object regions only). A single feature extractor is used instead of using two separate models with the almost same weight. Figure3 shows the unified structure of FR-CNN and RPN.Anchor boxes are a major part of modern object detectors. In object detection, a rectangular box is obtained for each object in the image, thus there are multiple boxes of various shapes and sizes in each image.

Figure2: FR-CNN Network with RPN

(7)

First, the images are divided into grids. The reason is that medical images are usually pretty big. And you want to make sure to have labels (cancerous or non-cancerous) that are done by professionals for each grid.

And each grid will be sent to the convolutional neural network to train. When sending each grid, a mask which belongs to the grid is also sent which says either “cancerous” or “non-cancerous”. Then, slide through each grid and make the neural network learn each grid with its mask.

k- Anchor boxes are generated for every pixel in the feature map (output of CNN). Thus the total number of anchor boxes is h*w*k(h*w is the output size of the feature map). K here is a hyperparameter. These k anchor boxes are of varying in sizes and aspect ratios which help covers objects of various shapes and sizes.

The anchor boxes are hyperparameter specific to dataset and type of objects in that dataset(For example in some medical data if the object can be only of single size, only 3 anchor boxes of different aspect ratios are required).

Instead of getting a raw regression output of an object, it is calculated as an offset to the anchor box.

This offset in most cases will be a slight shift of the anchor box as these anchor boxes are placed all over the image. An object that is detected will be overlapped by multiple anchor boxes and not just 1. These redundant predictions are then later removed using non-max suppression. The output of our model is of 4*k*h*w dimension(one box prediction for each anchor box, classification score is also predicted for each anchor box giving a probability of it containing object). This theoretically limits the number of objects that can be detected by CNN to 4*k*h*w, but in practice, this number is big enough.Anchor boxes can solve the issues of using multi scales at test time by using anchor boxes of various sizes(Red, Green, and Blue boxes in the above figure 4).

3.2 Region proposal network

Input of RPNis the feature map which produces a set of rectangular object proposals as output, each having an objectness score. The objectness measures score between object and background(thus low score for background and higher for regions with object). The region proposal time with selective search is 2 sec per image, whereas with RPN it‟s just 10ms.

FR-CNN uses anchor boxes of 3 aspect ratios and 3 scales. Thus, for each pixel in the feature map, there are 9 anchor boxes.The architecture is a simple convolution layer with kernel size 3*3 followed by two fully connected (FC) layers(objectness score(classification) and regression of proposals). This fully connected layer is implemented using 1*1 convolutional layers. The output size for the classification layer should be 2*9 (foreground and background) and 4*9 for the regression layer(Here 9 is a number of anchors for each pixel).

The total number of predictions is now intuitive and will be (4+2)*9*(H*W), for each pixel in feature map.

(8)

Figure: 3 Faster RCNN Loss Function

The loss function used in FR-CNN is 𝐿 𝑝_𝑖 , 𝑡_𝑖 = 1

𝑁_𝑐𝑙𝑠 𝐿_𝑐𝑙𝑠 (

𝑖

𝑝_𝑖, 𝑝_𝑖^∗) + 𝜆 1

𝐿_𝑟𝑒𝑔 𝑝_𝑖^∗𝐿_𝑐𝑙𝑠 (𝑝_𝑖, 𝑝_𝑖^∗) [1]

𝑖

In this equation, pᵢ represents predicted probability(output from cls), and pᵢ* is the ground truth similarly, tᵢ representspredicted bounding box and tᵢ* stands for ground truth bounding box. 𝐿𝑐𝑙𝑠 are classification loss(log loss) and Lreg is smooth L₁ loss.

As discussed earlier regression offset is calculated from the nearest anchor box. To relate it with region proposal technique, anchor boxes are now acting as region proposal. For a feature map of size 40*60, there are a total of 40*60*9 ~ 20000 anchor boxes. All anchor boxes don‟t contribute to loss at the training time. The anchors with largest IOU with ground truth and anchor with IOU overlap larger than 0.7 are given „+ve‟ labels.

Anchors with IOU less than 0.3 are labeled as „-ve‟. Anchors that are neither „+ve‟ nor„-ve‟has no contribution towards the training objective. Cross-boundary anchors are also ignored.

Training

FR-CNN architecture is a unified network with RPN and CNN layers are shared by both architectures.

RPN and Fast RCNN cannot be trained separately(it will give different weights and thus CNN has to be passed twice for each of them). The authors used a 4-step training algorithm discussed below:

(9)

1. RPN trained initially and pre-trained weights from image-net are used.

2. FR-CNN is trained with proposals of step-1.

3. RPN is trained using convolution layers from step 2 and only layers unique to RPN are updated(weights of convolution layer is not updated).

4. Keeping this convolution layer, layers unique to Fast R-CNN are fine-tuned.

At training time, 20000 anchors(proposals) which are the output of RPN are first reduced by removing cross-boundary anchors(giving 6000 anchors). Non-max suppression(NMS) is applied to remove redundant predictions with a threshold of 0.7(giving 2000 anchors). After NMS top-N ranked (classification score) proposal regions are used. For classification, Cross Entropy Losswith values 0 to 1 is used to measure theperformance.

Smooth L1 lossfor B-box regressionis used, an absolute value between ground truthand prediction as L1 loss is less sensitive to outliers than loss like L2which squares the error.

Testing

FR-CNN is a completely deep learning-based approach with a unified network and does not depend on algorithms like the selective search for proposals. Thus the image is directly passed to network giving predictions as output.

Every point in 37x50 is assumed as an anchor. Particular ratios and sizes are defined for every anchor (3 ratios and 3 sizes are 1:1, 1:2, 2:1 and 128², 256², 512² respectively forthe actual image).RPN is connected with the convolutional layer containing 3x3 filters, 1 padding, and 512 output channels. Output is associated with two 1x1 convolutional layer for classification (determines whether box is an object or not)and box- regression as in figure 6.

Figure4: Implementation ofConvolutional RPN architecture for k anchors

(10)

Here, each anchor has 9 boxes related to the actual image with the meaning that 16650 boxes exists in the actual image. But only 256 out of 16650 boxes are selected as a small group containing 128 foregrounds („+ve‟s) and 128 backgrounds („-ve‟s). Simultaneously, non-maximum suppression is applied to avoid the overlapping of the proposed regions ROI.RPN ends after the processing the above steps. The next stage of FR- CNN is ROI pooling which is used for ROI whose output is 7x7x512. This layer is then flattened with few FC layers. Finally, softmax function is applied for classification and linear regression for fixing the location of the boxes. The architecture of R-CNN is as in figure 7.

Figure.5: Architecture of R-CNN Flow chart of Proposed FR-CNN system:

Figure 8 illustrates the flow of the method introduced in this research work.

Figure.6: Overall flowchart for the proposed architecture

(11)

Preparing training data and labels (get_anchor_gt)

Annotation.txt is a file that produces Input containing a several image with the information of bounding boxes. RPN method is used for creating these proposed boxes.

Calculation of RPN for every image (calc_rpn)

In casethe size and anchor shapes of feature map are 18x25=450 and 9 respectively, then the potential anchors are 450x9=4050. Initially, the every anchorstatusis „-ve‟. Next, anchor is fixed as „+ve‟ in case IOU is more than 0.7. When IOU lies between 0.3 to 0.7, this one isuncertain and thus isn‟tconsidered in the intention.

The drawback is the RPN holds several „-ve‟ regions rather than „+ve‟, hence few „-ve‟ regions are turned off.Total „+ve‟ and „-ve‟ regions are restricted to 256.y_is_box_valid checks whether anchor has an object. y_rpn_overlap finds out whether this anchor and ground-truth bounding box overlaps.

Both y_is_box_valid and y_rpn_overlap are 1, If anchor is „+ve‟.

Both y_is_box_valid and y_rpn_overlap are 0, if anchor is neutral.

y_is_box_valid is 1 and y_rpn_overlap =0 for „-ve‟ anchor.

(12)

Calculation of ROI from RPN (rpn_to_roi)

 Arguments taken (num_anchors = 9)

rpn_layer: output layer for rpn classification

shape (1, feature_map.height, feature_map.width, num_anchors) is (1, 18, 25, 9)

regr_layer: output layer for rpn regression

shape (1, feature_map.height, feature_map.width, num_anchors*4)

(13)

is (1, 18, 25, 36) C: config

use_regr: checks if bboxes regression can be used in rpn max_boxes: maximumnumber of bboxes for NMS overlap_thresh: When iou> threshold in NMS, drop box

 Values Returned

result: boxes from NMS (shape is 300, 4) boxes: bboxes‟ coordinates

From the above step, for 4050 anchors, max_boxes have to be extracted as ROI and has to be passed to the classifier layer. In the function, at first, boxes overstepping original image are deleted. Then, NMS with the threshold value of 0.7 is used.

RoIPooling layer and Classifier layer (RoIPoolingConv, classifierlayer):

RoIPooling layer processesRoIwith a specific size with the use max pooling. Every RoI is partitioned as few subcells, and max pooling is put into to every sub-cell. The shape of the output is the number of subcells. Classifier layer, final layer lying behind RoIPooling layer is used for predicting the name of class for every anchor input and bounding box regression.

Initially,flatten the pooling layerfollowed with two FC layer and 0.5 dropout. At last, outputs two layers.

# out_class: softmax activation function which classifies class name of the object

# out_regr: linear activation function for coordinates ofbboxes regression

(14)

 Pooling layerof ROI, extraordinarytype of spatial pyramid pooling (SPP) layer, has a single pyramid level which primarily partitions the proposal windows selected obtains features that are the outcomes of algorithm of region proposal to sub windows with size h/Hbyw/W. Pooling operation is performed on every sub-window which produces output as fixed-size features (H x W) regardless of input size. The values chosen for H and W were7 like the output suits well with 1st fully-connected layer. ROI pooling is performed on each channel individually.

 ROI Pooling layer produces the output as N x 7 x 7 x 512 where N is the total proposals which are provided as inputs for the successive fully connected layers, softmax and BB-regression branches. Softmax classification branch gives values of probability of every ROI with K and one catch-all background category. Output of BB regression makes the bounding boxes from RPA more accurate.

Loss:

The softmax layer classification provides each ROI probabilities over (K +1) categories p = p₀, … pk. Loss of ClassificationLloc(p,u) is provided by -log(pᵤ) that is log loss on behalf of true class u.L1Smooth loss for BB regression is given by,

[2]

Regression gives 4 bounding box offsets tᵏᵢ where i = x, y, w, and h. (x, y) represents top-left corner while w and h are width and height of the bounding box. The true bounding box regression aiming at class u is given as vᵢ when u≥1. When u=0, it is ignored sincebackground classes have no groundtruth boxes. Smooth L loss is given by equation 1 and multi-task loss of every ROI is the integration of two losses given by equation 2. FR-CNN having anintegratedlearning scheme, fine-tunes backbone CNNandperforms classification and regression of bounding box.

Training:

 Each mini batch with 64 ROIs is formed,from N=2 images at the training stage. Similar to RCNN, ROI having twenty five percentage are proposals of object with IoU>0.5with a ground-truth bounding box of a foreground class which is „+ve‟and labeled with suitable𝑢 = 1 … 𝐾.

 The remainingROIs are sampled from proposals withIoU having groundtruth boxes bounded between [0.1, 0.5) are labeledas background class 𝑢 = 0.

(15)

 Prior to pooling layer ROI, the whole image passes throughoutCNN. Hence, every ROI of the identicalimage dividesmemory and computation in both frontwardas well astoward the backgo bythroughout CNN.

 The sampling of every ROI of very few imagesgoes through feature generation, classification, and regression modules.But, in SPP Nets, almost all ROIs of different images foes through training process to detect only finetunes the completely connected layers after the feature productionas it is infeasible for updating weights ahead of the layer SPP. Few ROI‟s have huge receptive fields on the actual image.

Original features at a haltare extracted from a network which is pre-trained purposefully trained for classification that restricts the SPP Nets accuracy.

 The batch size of the networkis too tiny for Convolutional Neural Networkuntilthe batch size of ROI pooling layer is 2, but larger for softmax and regression layers with batch size 128.

 Back propagation through ROI pooling layer: Every mini-batch ROI r, consider ROI pooling output unit yᵣⱼ as max-pooling output of sub-window R(r, j). After then, gradient is accumulated in the input unit (xᵢ) in R(r, j) when position i is the argmax chosen for yᵣⱼas in figure 7.

Figure.7: Back propagation through ROI pooling layer

 In FR-CNN multi-scale pipeline, inputs are resized as randomly sampled size during training such that it is scale invariance. During testing, every image is given to the network with multiple fixed scales. For every ROI, features are pooled from any one of these chosen scales such that the pixels of scaled candidate window are closer to 224 x 224. But,single scale pipeline performs better withless computing time cost. In the single-scale method, every image during training and testing are resized to 600 with upper cap of 1000.

(16)

 Larger fully connected layers are compressed using truncated SVD approach such that network is more effective. Then, a layer parameterized by W, weight matrix, is factorized to minimize parameter count by dividing it as two layers (ΣtVᵀ and U with biases) withno non-linearity between these layers, where W ~ U ΣtVᵀ.

SVC (Support VectorClassifier):

SVCs, a kind of large-margin classifier, are a set of supervised learning approaches which learns the dataset utilized for both classification and regression. Moreover, the aim of SVC vector space-based machine learning approach is to discover decision boundary between two classes which is far from any point in the training data.SVC classifies by generating an abstraction of vectors named as support vectorswhich is a part of training vectorset. Data can be separated as training and testing set to achieve better classification. In training set, for every occurrence, there exists a target value and several features. SVC is employed as it provides a model based on training data. Using the features of the test data, target values are forecasted. SVCs primary objective is to classify tasks but it has been extended to learning tasks and regression. SVC, generally a binary classifier, produces either „+ve‟ or „-ve‟ output of the learning function. The linear boundary of the data is used to classify patterns between two classes.Support Vectors are the co-ordinates of individual observation.Kernel approach is employed in SVCfor non-linear classification which converts low dimensional input space to higher ones. Moreover, non-separable classes are converted as separable ones based on data labels.

When SV points are far from hyperplane, the probability of classifying the points correctly is also more in their corresponding region. SV points critically determines hyperplane as the change in the position of vectors alters the position of the hyperplane. This hyperplane is technically known as margin maximizing hyperplane.

This algorithm aims in finding a hyper plane with N-dimensions where the data points are classified such that the margin is maximizd. N dimension varies as per the number of features. Comparison of two features is carried out smoothly. However, when numerous features are classified, it is not straightforwardalways. When the margins are maximized, more accurate prediction is obtained.

Gaussian NB (NaivesBayes):

Naïve Bayes is a fast and straightforward approach used for classification which works on the basis of Bayes theoremand is represented as:

𝑃 𝑋 𝑌 =𝑃 𝑌 𝑋 𝑃(𝑋) 𝑃(𝑌) [3]

(17)

This classifier considers that every variable equally contributes to the outcome independently. At this point, every feature is independent to each other and output is also affected with same weight. Hence, Naïve Bayes theorem cannot be applied to the issues related to real-lifeand only low accuracy is obtained when this algorithm is used. Hence, Gaussian NB, a kind of NB, is used which considers that the features adopt normal distribution. The possibility of features is assumed to be Gaussian and has a conditional probability. Gaussian NB is given below:

𝑃 𝑥𝑖 𝑦 = 1

2𝜋𝜎2𝑦exp −(𝑥𝑖 − 𝜇𝑦)²

2𝜎² [4]

4. EXPERIMENTAL RESULTS AND PERFORMANCE ANALYSIS

In this article, the proposal for a CNN-based approach for classifying ovary tumors in MRI images.

Figure 8 Selecting ensemble Classification Algorithm to detect ovarian cancer

Figure 9 Identification of Ovarian Tumor Images

This section discusses about the performance analysis for the proposed technique in ovarian cancer detection which is compared with the existing methods.Confusion matrix shows the model performance based on true „-ve‟s, true „+ve‟s, and false„-ve‟ and false „+ve‟.

(18)

Figure.10: Confusion matrix ofGaussian NB using FR-CNN

The above figure 10 shows the confusion matrix of Gaussian NB using FR-CNN in which the rows represent the predictedclass (output class) and columns denotes the actual class (target class) of data pertaining to ovarian cancer. The diagonal blueand white cells denote the trained network that are correctly and incorrectly classified.The column on the right side indicates every predicted class while the rowat bottom represents the performance of every actual class. This confusion matrix plot for Gaussian NB using FR- CNNshows that theoverall classification attains 98.69% correct classificationperformance.

Figure.11: Confusion matrix of SVC using FR-CNN

(19)

The above figure 11 shows confusion matrix of SVC using FR-CNN with rows and columns indicates predicted and actual classes respectively.This confusion matrix plot for SVC using FR-CNN shows that all overall classification achieves 97.39%. The column on the right side indicates predicted class while the row at bottom represents the performance of every actual class.Here, zeroes are hidden in order to easily analyze the performance. From this confusion matrix, more often few couples are recognized wrongly. Table 1 shows the analysis of SVC and Gaussian NB with various parameters.

Table1: Analysis of SVC and Gaussian NB Classifiers

Parameters SVC (%) Gaussian NB (%)

Precision 95.96 97.7

Recall/Sensitivity 94.31 97.7

Specificity 97.39 98.69

Figure.12:Classification graph of Gaussian NB and SVM using FR-CNN

The above figure 12 shows the graphical representation based on the parameters for Gaussian NB and SVM using FR-CNN. The parameters taken are precision, Recall/sensitivity and specificity has been calculated in %. In precision value of SVC gives 95.96%, and Gaussian NB gives 97.7% where precision is enhanced in Gaussian NB using FR-CNN. For recall/sensitivity it gives 94.31% for SVC and 97.7% using Gaussian NB and also for specificity SVC value is 97.39% and 98.69% for Gaussian NB using FR-CNN.

From this above discussion, the Gaussian NB technique in classification using FR-CNN gives enhanced

(20)

predicted class in ovarian cancer detection. Table 2 compares the parametric values obtained for various methods.

Table2: Comparative Analysis of various parameters for various methods

Parameters KNN (%) CNN (%) DCNN (%) SVC (%) Gaussian NB (%)

Precision 78.45 81.91 89.19 95.96 97.7

Recall/Sensitivity 74.19 79.02 88.28 94.31 97.7

Specificity 95.33 82.93 91.91 97.39 98.69

Figure.13: Overall comparison between proposed and existing algorithms

The above graph shows overall comparison for precision, recall, specificityfor existing and proposed techniques. For precision, CNN gives 81.91%, DCNN gives 89.19%, KNN gives 78.45%, SVC gives 95.96%

and Gaussian NB gives 97.7% among these techniques SVM and Gaussian NB gives optimum value. For recall, CNN is 79.02%, DCNN is 88.28%, KNN gives 74.19%, SVS gives 94.31%, and Gaussian NB gives 97.7%. For specificity, CNN gives 82.93%, DCNN gives 91.91%, KNN gives 75.33%, and SVC gives 97.39% and Gaussian NB 98.69%. Among all the techniques proposed SVC and Gaussian NB using FR- CNNhas obtainedenhanced value.

(21)

5. CONCLUSION

Hence from the performance analysis, it shows that the classification technique of both SVC and Gaussian NB using FR-CNN gives the precision value more than 95% when compared with existing techniques. Among this classification technique 97% to nearly 99% of precision has obtained from the predicted class using this proposed FR-CNN. Based on the performance of the proposed model, it is concluded that this ovarian cancer detection classification model is a vital contribution in the medical field which helps the physicians to take the precise decision and treat the patients in a better way. However, no public ovarian cancer dataset was considered for classification in this research. Hence, it can be focuses to use ovarian cancer dataset for classification as the future enhancement of this research work.

REFERENCES

[1] Landrum LM, Java J, Mathews CA, Lanneau Jr GS, Copeland LJ, Armstrong DK, “Prognostic factors for stage III epithelial ovarian cancer treated within traperitoneal chemotherapy: a Gynecologic Oncology Group study”,Gynecol Oncol, vol.130, no.1, pp.12–8, 2013.

[2] Mangone L, Mandato VD, Gandolfi R, Tromellini C, Abrate M,“The impact of epithelial ovarian cancer diagnosis on women‟s life: a qualitative study”, European Journal of Gynaecol Oncol, vol.35, no.1, pp.32–8, 2014.

[3] BHP,“Health Survey by the Bureau of Health Promotion

(BHP)”,https://cris.hpa.gov.tw/pagepub/Home.aspx, 2017.

[4] Nolen BM andLokshin AE,“Screening for ovarian cancer: old tools, new lessons”,Cancer Biomark, vol.8, no.4, pp.177–86, 2010.

[5] Liang C and Peng L,“An automated diagnosis system of liver disease using artificial immune and genetic algorithms”,Journal of medical systems, vol.37, no.2, pp.1-10, 2013.

[6] Jelovac D and Armstrong DK,“Recent progress in the diagnosis and treatment of ovarian cancer”,CA Cancer Journal of Clin, vol.61, no.3, pp.183–203, 2011.

[7] Karbalay-Doust S and Noorafshan A, “Stereological estimation of ovarian oocyte volume, surface area and number: application on mice treated with nandrolone decanoate”,Folia Histochem. Cytobiol, vol.50,pp.275–279, 2012.

[8] İnik and Özkan,“A new method for automatic counting of ovarian follicles on whole slide histological images based on convolutional neural network”, Computers in biology and medicine, vol.112,2019.

[9] Massimo Mazzini1 an Franco Giorgi, “The follicle cell-oocyte interaction in ovarian follicles of the stick insect Bacillus rossius (Rossi): (Insecta: Phasmatodea)”,Journal of Morphology, vol.185, no.1, pp.37-49,2016.

[10] González-Reyes A and St Johnston D, “Patterning of the follicle cell epithelium along the anteriorposterior axis during Drosophila oogenesis”, Development, vol.125,no.15, pp.2837-2846, 1998.

(22)

[11] Drosophila oogenesis, “The Wellcome/CRC Institute and Department of Genetics”, University of Cambridge, pp.2837-2846,2018.

[12] Carole L. Browne and William Werner, “Intercellular junctions between the follicle cells and oocytes ofXenopus laevis”,Journal of Experimental Zoology, vol.230, no.1, pp.105-113,2014.

[13] Wu and Miao,“Automatic classification of ovarian cancer types from cytological images using deep convolutional neural networks,” Bioscience reports, vol.38, no.3,2018.

[14] Shafi, Uroosa and Sugandha Sharma,“Ovarian Cancer Detection in MRI Images using Feature Space and Classification Method”, International Journal of Recent Technology and Engineering (IJRTE),vol.8, no.2,2019.

[15] Vasavi G and Jyothi S,“Classification and detection of ovarian cysts in ultrasound images”,International Conference on Trends in Electronics and Informatics (ICEI),pp.783-787, 2017.

[16] SchorgeJ. O, Modesitt S. C, Coleman R. L, Cohn D. E, Kauff N. D, Duska L. R and Herzog T. J,

“SGO White Paper on ovarian cancer: etiology, screening and surveillance”,Gynecologic oncology, vol.119, no.1, pp.7-17, 2010.

[17] Hiremath P. S and Jyothi R. Tegnoor, “Automated ovarian classification in digital ultrasound images”, International Journal of Biomedical Engineering and Technology, vol.11, no.1, pp.46-65, 2013.

[18] Nawgaje, Devesh D and Rajendra D. Kanphade, “Hardware Implementation of Genetic Algorithm for Ovarian Cancer Image Segmentation”, Proceedings of the International Journal of Soft Computing and Engineering (IJSCE), vol.2, no.6, pp.304-306, 2012.