Classification of stages in cervical cancer MRI by customized CNN and transfer learning

Abstract

Cervical cancer is the common cancer among women, where early-stage diagnoses of cervical cancer lead to recovery from the deadly cervical cancer. Correct cervical cancer staging is predominant to decide the treatment. Hence, cervical cancer staging is an important problem in designing automatic detection and diagnosing applications of the medical field. Convolutional Neural Networks (CNNs) often plays a greater role in object identification and classification. The performance of CNN in medical image classification can already compete with radiologists. In this paper, we planned to build a deep Capsule Network (CapsNet) for medical image classification that can achieve high accuracy using cervical cancer Magnetic Resonance (MR) images. In this study, a customized deep CNN model is developed using CapsNet to automatically predict the cervical cancer from MR images. In CapsNet, each layer receives input from all preceding layers, which helps to classify the features. The hyper parameters are estimated and it controls the backpropagation gradient at the initial learning. To improve the CapsNet performance, residual blocks are included between dense layers. Training and testing are performed with around 12,771 T2-weighted MR images of the TCGA-CESC dataset publicly available for research work. The results show that the accuracy of Customized CNN using CapsNetis higher and behaves well in classifying the cervical cancer. Thus, it is evident that CNN models can be used in developing automatic image analysis tools in the medical field.

Introduction

Cervical Cancer is a prevalent cancer worldwide after colorectal, breast and lung cancers, where this cancer affects predominantly the women all over the world (Ghoneim et al. 2020). As a result, people are not as knowledgeable about the disease, and as a result, lack of awareness and inadequate access to the healthcare related services is a greater challenge. In contrast, developed countries have screening strategies that ensure the ability to reliably and effectively screen for precancerous lesions, which allows for detection and treatment at an earlier stage. According to the strategy of reducing the prevalence of cervical cancer, the best method is to screen and vaccinate in order to lead healthy lives and safeguard all (Deng et al. 2018) (Zhu et al. 2019).

However, the diagnosis at early stages is difficult since the cancer develops without any symptoms and it appears at the later stages that poses the risk of spreading across all regions. Hence, it is very necessary to screen at earlier stages to increase the rate of survival.

Cervical Screening is an effective programmes that tends to reduce the morbidity and mortality associated with the disease (Vijayakumar and Vinothkanna 2020). The successfulness of screening depends on various factors that includes ability to find facilities, the quality of screening tests, and the adequacy of follow-up in addition to early diagnosis and treatment of any lesions found (Li et al. 2020). Screening services for cervical center are underfunded in low and middle-income countries because only a few skilled health workers are available, and the healthcare resources required to maintain screening programmes cannot be sustained (Mobiny et al. 2019) (Pandian 2019).

In recent past, Artificial Intelligence in association with machine learning plays a major part in classification of disease instances in healthcare industry. The ratio of disease prevalence gets increasing and hence machine learning resolves the optimization problem in classification process associated with disease diagnosis.

This study aims to adapt the Convolutional Neural Network (CNN) network to classify medical images in order to predict stages of cervical cancer. CNN has been demonstrated extremely fruitful in image classification problems. Fundamentally CNN improved performance in classifying many image databases such as MNIST and CIFAR10. It has been proving in learning the structures from an image. Most objects with such structures constructs complex features from local features including curves and edges.

Recently, CapsNet has also been incorporated into medical imaging analysis, such as brain tumor segmentation (Goceri 2020), cancer screening (Mobiny and Nguyen 2018), diagnosis (Yadav and Jadhav 2019), segmentation (Bonheur et al. 2019), and many more. The conventional models observe that medical images can be classified using CapsNet and perform well on such data. Now a days deep networks are popular and achieving high accuracy in image classification and image recognition. Adding many layers with different combinations has come up with significant improvement in performance.

In this paper, a customized CNN-CapsNet is built to adapt a multi-class image classification problem and resolves the classification problem associated with cervical cancer MR images. The paper is organized as follows. Literature survey is briefed in Sect. Two . The methodology ant training of the model is described in Sect. Three. Section Four provides the evaluation of CNN-CapsNet in case of classification. The conclusion is given in Sect. five

Literature survey

Medical image classification of the cancerous and normal cells is a very important process in deciding the right treatment at the right time. Many publications are devoted to the classification of an object using CapsNet in the literature.

Xiang et al. (2021) offers a CNN-based tumor classification with a 3-D Res-CapsNet model, where the input image dataset is a breast ultrasound one. This model uses residual block and group normalization in order to achieve higher accuracy.

Lin et al. (2021) proposed a CapsNet with scale invariant feature transform for boosting representability. The hyperparameter setting, customized structure, and dynamic routing are discussed for the purpose of classification.

A methodology called CapsNets was created by the group of researchers consisting of Deng et al. (2018) in order to increase the classification accuracy with limited samples. This analysis is of further assistance in understanding CNN and CapsNet architectures as similar paradigms of network design. Following this, the data was scrutinised and debated in the presence of uncertainty analysis and probability maps.

Conv-Capsule was developed by the Zhu et al. (2019) group who added full, local connections with transform matrices. While training a convolutional network tends to exaggerate its network performance, an earlier study suggests that the reduction in trainable parameters in Conv-Capsule could mitigate overfitting with available training samples.

To ensure the classification quality of the font styles, Vijayakumar and Vinothkanna (2020) imposed the CapsNet algorithm. This proposed method has been found to be accurate when it classifies fonts. The alphabet and character spacing and the results of the performance evaluation, which looks at these factors are made public.

In the research published by Li et al. (2020), they described CapsNet on hyper-spectral image (HSI) classification with maximum correntropy that reduces the outliers and noise. This results in a more robust and generalizable model. CapsNet framework relies on a highly robust, expanded version of CapsNet, is also proposed for data fusion and classification. HSI is taken from three well-established hyperspectral datasets and used to show how our deep learning models outperform the competition.

Using a new algorithm, researchers at Mediomobiny et al. (2019) designed a variant of CapsNets that is faster and simpler than CNNs. By employing the information found in videos of microscopy samples, we propose a recurrent CapsNet, which is made up of a bi-directional long short-term recurrent structure stacked on top of a CapsNet. The experiments show that, given the constraints of time, the CapsNet performs at 93.8% accuracy and prvides consistent performance than CNN.

Pandan (2019) offered a helpful methodology for classifying and identifying cancer cells via images of pathological tissue, with the aid of the capsule network. The CapsNet learns the patterns from training data that can be utilized for particular applications. The early detection of cancer is aided by this, and locating the root cause is necessary for curing the disease. Doing so will help progress the cancer into a state where it can no longer be active.

In order to identify irises, researchers studied the capsule network architecture, and devised a deep learning method called CapsNet based on it. To ensure that the network is functioning correctly, we make structural adjustments to the network and implement a routing between the capsules for optimal adaptation to iris recognition. In the event that the number of samples is constrained, we make it possible to use the deep learning method. For the three networks, we divide the overall network structure into a subnetwork based on constituent blocks.

To adapt CAD models to new domains, a CapsNet was proposed by the aforementioned researchers, who were led by Mobiny, A., et al. A feature embedding extraction network connects a number of capsule nets that all serve the same function, extracting feature embeddings from target domain by incorporating knowledge learned during training into a set of artificial memories. With smaller annotated samples, CapsNet can efficiently adapt to all kinds of domains.

This shows not only the need for automation of classification of normal and abnormal images but also shows that research in that area is still in progress. Many methods (Zhao et al. 2019; Mobiny et al. 2021) are proposed in the literature to automate the classification of cancer from medical images. Recent works on the automatic classification of medical images provide performances due to the advanced development of deep learning concepts.

Methodology

Classification of cervical cancer into different stages is a challenge for radiologists due to within-class variation and between-class similarity of contrast in MR images. To assist radiologists, we propose a CNN based cancer classification engine using customized CNN model built with CapsNet to classify MR images of cervical cancer. The workflow of the proposed system is shown in Fig. 1. The customized CNN with CapsNet model is fine-tuned to produce good results.

Fig. 1.

Workflow of proposed work

Preprocessing technique

The preprocessing technique involves noise removal from the MR images and contrast enhancement for optimal classification of objects in cervical cancer images.

Noise removal

Noise is unwanted information present in digital images. Images without noise are vital for further diagnosis process. Removing noise from the original MR images is a challenging problem. This noise can affect the images in many ways, such as blurring, artifacts, etc. The adaptive median filter effect is given in Fig. 2. Depending on the input data source, the noise present in an image is classified into different categories. The utilization of filters suppress the presence of noise during the acquisition of images. In image processing, filters enhance the images by reducing unwanted frequencies and by smoothening an image. Various types of filters exist, like mean filters, adaptive filters, etc. Without affecting the original image, adaptive filters reduce noise. During filtering, the statistical parameters are estimated and the values are adapted with the modification of pixel values. In our work, we use an adaptive median filter to remove noise. Calculate the median, minimum, and maximum pixel value of filter size. Then compares each pixel value to either replace the pixel value or keep the pixel value. Then increase the filter window size. The adaptive median filter only affects image pixels determined to have noise content. It performs well for low and high density of noises.

Fig. 2.

Adaptive median filter output a Original image b After filtering

Contrast enhancement

In medical imaging, the quality of the image is vital to diagnose diseases. The challenges of staging of cancer lie in the intensities of tissues. Figure 3 shows Intensity Enhancement output using Adaptive Histogram Equalization. Contrast enhancement techniques improve the intensity variation of images that helps to detect the cancerous tissue. Contrast enhancement helps in finding the features associated with the tissue heterogeneity. Further, it allows the CNN model to generate features for classifying the MR images. In this research work, we use contrast enhancement on images using Adaptive Histogram Equalization that improves the MR image intensity.

Fig. 3.

The output of Intensity Enhancement a Before enhancement b After enhancement

CapsNet

The study presents the CapsNet architecture illustrated in Fig. 4 that accurately classify the benign and malignant cancer cells in images.

Fig. 4.

Customized CNN-CapsNet model

Capsules in the CapsNet is regarded as a group of neurons that represents the activity of parameters of neurons. The vector length (4 × 4) in ReLU Conv1 layer indicates that the specific entity or the features are likely existing. The CapsNet overcome the weakness of CNN that are linked to pooling layers and suitable modification of pooling layers in CapsNet improved the classification criteria. However, the coupling coefficients may vary from the parent capsules and it may vary with increasing iterations. The parent capsule hence classifies the event by increasing the coefficient between the capsules, where the prediction conforms the parent capsule output.

In view of u_i as capsule output i, where the parent capsule j is calculated for the process of classification:

$$u_{j|i}=W_{\mathit{ij}}{u}_i$$ 1

whereu_j|i- vector prediction obtained from the output of capsulej, where the capsule iestimates the vector prediction, and.W_ij- weighting matrix applied on backward pass

Softmax function used in CapsNet for coupling coefficients c_ij using the agreement level between the capsule in the subsequent layer:

$$c_{\mathit{ij}}=\frac{exp\left(b_{\mathit{ij}}\right)}{\sum _{k}exp\left(b_{\mathit{ij}}\right)}$$ 2

whereb_ij-log likelihood function, where it functions if the ith capsule is interconnected with jth capsule and it is set to 0 at the initial iteration.

The input capsule vector j is estimated as below:

$$s_j=\sum _i{c}_{\mathit{ij}}{u}_{j|i}$$ 3

The study uses a non-linear function that exceeds each capsules output vector and then uses initial vector to obtain the output, where its non-linear function is defined as below:

$$v_j=\frac{{∥s_j∥}^2}{1+{∥s_j∥}^2}\frac{s_j}{∥s_j∥}$$ 4

wheres_j-input of jth capsule andvj-output of jth capsule

The log probability during the process of classification is updated in terms of agreement between the output capsule vector using non-linear function v_j and output capsule vector without non-linear function u_j, where if the vectors tends to agree and a large internal product is then provided.

Hence to update the coupling coefficient in pooling layer and log probability, the study uses an agreement function and this is estimated as below:

$$a_{\mathit{ij}}=v_j·u_{j|i}$$ 5

The last layer or RouteCaps with capsule k is designed with a loss function l_k, where it offers a higher level of accuracy in predicting the instances. The loss function l_k is further estimated as below:

$$l_k=\lambda \left(1-T_k\right)max{\left(0,\left|\left|v_k\right|\right|-m^{-}\right)}^2+T_{k}max{\left(0,m^{+}-||v_k||\right)}^2$$ 6

Here

$$T_k=\left\{\begin{array}{cc}1& \text{if}k\text{exist}\\ 0& \text{if}k\text{does}\text{not}\text{exist}\end{array}\right)$$

The hyper parameters including (m⁺, m⁻ and λ) during the process of training is hence estimated, where m +  = 0.9, m −  = 0.1, and λ = 0.5 is set prior training and λ controls the backpropagation gradient at the initial learning. Hence, the true labels are highly obtained with three layers in the CapsNet that includes a convolutional layer and two capsule layers.

CapsNet model tends to use reconstruction loss model in the form of a regularization function that tends to reduce the problems associated with over-fitting during the training of the CapsNet parameters. The capsules are hence allowed to encode the input as much as possible and hence the reconstruction is carried out through the feeding of 16D output over its capsules (final layer) to a neural network.

Results

In this section, a comparitive analysis between the proposed CapsNet model and conventional CNN models to classify the cervical cancer images. The subsection that follows shows the evaluation on various images.

Dataset

In our work, we have utilized MR images of the cervical cancer dataset downloaded from TCGA-CESC (Fig. 5).

Fig. 5.

Images used for classification of cervical cancer belonging to various stages

It is freely available for educational and research purposes. It has around 19,135 MR images of 54 patients. MR images in various planes, such as sagittal, axial, and coronal, are accessible. Both T1 and T2 type images are present in the dataset. Table. 1. Shows the distribution of stages of cervical cancer MR images found in the dataset.

Table 1.

Four stage of cervical cancer dataset

Stage	No. of patients	No. of images
Stage I	13	4480
Stage II	21	7649
Stage III	8	2219
Stage IV	12	4787

Discussion

The proposed CNN-CapsNet for the classification of cervical cancer is compared with conventional VGG-16 and CNN algorithm, where it helps in classification of various stages of cancer that includes stages I through IV based on the spread of cancer. The proposed method is emphasized on a fivefold cross validation model for the purpose of testing and the metrics are given below:

$$Accuracy=TP/(TP+TN+FP+FN)$$

$$Sensitivity=TP/(TP+FN)$$

$$\text{Specificity = TN/(TN + TP)}$$

$$\text{F - measure = TP/(TP + 0.5(FP + FN))}$$

$$MAPE=\frac{1}{n}\sum _{t=1}^n\left|\frac{A_t-F_t}{A_t}\right|$$

where n – Total iterations.

A_t – Actualvalue.

F_t – Forecastvalue.

TP shows the True Positive rate.FP shows the False Positiverate.TN shows the True Negativerate.

FN shows the False Negativerate.

Figure 6 illustrates the accuracy of the CNN-CapsNet classifier with conventional classifiers that includes VGG-16 and CNN. The CNN-CapsNet offers multi-label classification of input images and it follows supervised learning with higher number of training samples using five-fold cross validation. The testing shows that the proposed classifier achieves higher rate of classification accuracy than other models.

Fig. 6.

Accuracy

Figure 7 illustrates the F-measure of the CNN-CapsNet classifier with conventional classifiers that includes VGG-16 and CNN. The training on multiple images makes the CNN-CapsNet to achieve higher rate of F-measure in terms of higher specificity and sensitivity using five-fold cross validation. The testing shows that the proposed classifier achieves higher rate of F-measure than other models.

Fig. 7.

F-measure

Figure 8 illustrates the G-mean of the CNN-CapsNet classifier with conventional classifiers that includes VGG-16 and CNN. The repeated training on several input images makes the CNN-CapsNet to achieve G-mean rate with reduced errors using five-fold cross validation. The testing shows that the proposed classifier achieves higher rate of G-mean than other models. The higher G-mean mean to produce higher rate of classification with more positive classes associated with accurate classification than the existence of negative classes.

Fig. 8.

G-mean

Figure 9 illustrates the MAPE of the CNN-CapsNet classifier with conventional classifiers that includes VGG-16 and CNN. The repeated iteration of five-fold cross validation with a sigmoid loss function makes the CNN-CapsNet to accurately classify the classes of cervical cancer than other methods. The testing shows that the proposed classifier achieves reduced MAPE than other models.

Fig. 9.

MAPE

Figure 10 illustrates the sensitivity of the CNN-CapsNet classifier with conventional classifiers that includes VGG-16 and CNN. The CNN-CapsNet offers multi-label classification of input images allows the retrieval of more true positive instances than other instances in a five-fold cross validation. The testing shows that the proposed classifier achieves higher rate of sensitivity than other models.

Fig. 10.

Sensitivity

Figure 11 illustrates the specificity of the CNN-CapsNet classifier with conventional classifiers that includes VGG-16 and CNN. The CNN-CapsNet offers multi-label classification of input images allows the accurate classification of true negative instances in a five-fold cross validation. The testing shows that the proposed classifier achieves higher rate of specificity than other models.

Fig. 11.

Specificity

Finally, we executed the proposed customized CNN-CapsNet model using using NVIDIA GeForce GTX 1080 with 8 GB RAM system. Training, validation, and testing are taken place using 12,771 images. The MR images are in Dicom format. Dicom images are loaded using pydicom package of python and then passed as input to the customized CNN model, VGG16. The Table 2 provides the results of accuracy between the three models of CNN.

Table 2.

Overall classification accuracy

Model	Accuracy (%)
CNN-CapsNet	90.28
VGG 16	93.02
CNN	84.33

Conclusion

In this research work, we develop a customized CNN model with CapsNet for classifying MR images of cervical cancer. It produces an accuracy of 90.28% to classify eight stages of cervical cancer. Considering everything, many configurations of network architecture with different hyperparameters helped the customized CNN model to produce better classification accuracy. The study suggests that the CNN-CapsNet classification algorithm may improve the reliability of tools required for doctors to classify the medical images. Hence diagnosis accuracy improves, which helps for treatment planning in healthcare.In future, dense networks can be employed further to improve the testing with less data supplied to train the classifier.