Cervical cancer classification using convolutional neural networks and extreme learning machines

Highlights

• CNN-based cervical cancer detection and classification system is proposed.

• Extreme learning machine (ELM)-based classifier is used after the CNN.

• 99.7% accuracy is obtained in the 2-class problem using the Herlev database.

• 97.2% accuracy is obtained in the 7-class problem.

Abstract

Cervical cancer is one of the main reasons of death from cancer in women. The complication of this cancer can be limited if it is diagnosed and treated at an early stage. In this paper, we propose a cervical cancer cell detection and classification system based on convolutional neural networks (CNNs). The cell images are fed into a CNNs model to extract deep-learned features. Then, an extreme learning machine (ELM)-based classifier classifies the input images. CNNs model is used via transfer learning and fine tuning. Alternatives to the ELM, multi-layer perceptron (MLP) and autoencoder (AE)-based classifiers are also investigated. Experiments are performed using the Herlev database. The proposed CNN-ELM-based system achieved 99.5% accuracy in the detection problem (2-class) and 91.2% in the classification problem (7-class).

Introduction

The cervix of a human is covered by a thin layer of tissues consisting of cells. If a cell is changed into a malignant cell that can grow and divide rapidly and becomes a tumor, we call this situation the cervical cancer. This cancer can be treated if it is detected at an early stage. The diagnosis is normally done by a screening process and a biopsy. Image processing techniques can be applied to find the spread of the cancer. Cervical cancer is the fourth-most common cause of death from cancer in women [1].

Medical image processing and intelligent systems play a role in the analysis of the malignant cells. With the development of new techniques, they become cost-effective and less time-consuming. They are now becoming popular over conventional methods such as Pap Smear, Colposcopy, and Cervicography. These techniques are unbiased to human experience; however, we want to stress that they cannot replace the subjective (expert doctor) evaluation, but can assist them to a great degree.

State-of-the-art machine learning techniques and wireless communication technologies have enabled us to develop a complete medical diagnosis system that can operate in real-time, accurately, and without human interaction. Yet, there are many issues that need to be solved, for example, packet loss during transmission, high bandwidth requirement for medical video data transfer, and a robust algorithm to deal with many variations in data. To address some of this issues, edge-based cloud computing was proposed for voice pathology detection [2], [3], Internet of Things (IoTs) and cloud-based framework was realized in [4], deep learning for emotion recognition was proposed in [5], edge-based communication was introduced in [6], and a disease monitoring system was proposed in [7].

Computer-aided systems in cancerous cell detection have been used in the literature for quite a few times. In breast cancer recognition, different feature extraction methods such as local binary pattern, histogram of gradient orientation, and Laplacian Gaussian filter were used [8], [9], [10]. Local texture analysis was used to diagnose pulmonary nodules in [11]. To analyze dermoscopy images for skin cancer, directional filters and color component features were used in [12], [13]. A method for voice pathology detection using different input modalities were proposed in [14], [15].

Recently, deep learning has brought a big improvement in accuracy in many applications. Due to its high accuracies in many areas, it has become the state-of-the-art machine learning technique. A good survey on various cancerous cells detection using deep learning can be found in [16]. Deep learning was successfully used in EEG pathology detection [17], [18], environment classification [19], lung nodule detection [20], breast cancer detection [21], skin cancer detection [22], medical image analysis [23], audio–visual emotion recognition [24] and diseases prediction [25], [26]. With the increase in many types of sensors, processing of Big Data has added an extra dimension to the deep learning. Big Data has successfully handled in several medical related applications [27], [28].

Due to the success of deep learning in many medical applications, in this paper, we propose a deep learning-based system to detect and classify cervical cancerous cells. In particular, we use convolutional neural networks (CNNs) followed by an extreme learning machine (ELM)-based classifier in the system. We investigate different models of CNNs via transfer learning. We also investigate various classifiers such as multi-layer perceptron (MLP), ELM, and autoencoder (AE). A public database, the Herlev database [29], has been used in the experiments. The contributions of the work are (i) introducing CNNs in cervical cancer cell detection and classification, (ii) introducing ELM-based classifier, and (iii) introducing AE-based classifier. The CNN extracts deep-learned features from the raw input. These features are fed into an ELM-based classifier or an AE-based classifier. To the best of our knowledge, this is the first attempt to investigate the use of the ELM and the AE after the CNN in the cervical cancer detection system.

The paper is organized as follows. Some related previous works are briefed in Section 2. The proposed system is described in Section 3. Experimental results and discussion are provided in Section 4. The conclusion of the work and some future directions are given in Section 5.