Elsevier

Neural Networks

Volume 128, August 2020, Pages 47-60
Neural Networks

Automated classification of cells into multiple classes in epithelial tissue of oral squamous cell carcinoma using transfer learning and convolutional neural network

https://doi.org/10.1016/j.neunet.2020.05.003Get rights and content

Highlights

  • Study on biopsy images of more prevalent oral squamous cell carcinoma (OSCC).

  • A new CNN technique for multi-class cell type classification of OSCC is proposed.

  • Transfer learning approach with four candidate pre-trained models is also applied.

  • The proposed CNN model outperforms transfer learning approaches.

  • Classification accuracy achieved is 97.5%, the highest in comparison to date.

Abstract

The analysis of tissue of a tumor in the oral cavity is essential for the pathologist to ascertain its grading. Recent studies using biopsy images reveal computer-aided diagnosis for oral sub-mucous fibrosis (OSF) carried out using machine learning algorithms, but no research has yet been outlined for multi-class grading of oral squamous cell carcinoma (OSCC). Pertinently, with the advent of deep learning in digital imaging and computational aid in the diagnosis, multi-class classification of OSCC biopsy images can help in timely and effective prognosis and multi-modal treatment protocols for oral cancer patients, thus reducing the operational workload of pathologists while enhancing management of the disease. With this motivation, this study attempts to classify OSCC into its four classes as per the Broder’s system of histological grading. The study is conducted on oral biopsy images applying two methods: (i) through the application of transfer learning using pre-trained deep convolutional neural network (CNN) wherein four candidate pre-trained models, namely Alexnet, VGG-16, VGG-19 and Resnet-50, were chosen to find the most suitable model for our classification problem, and (ii) by a proposed CNN model. Although the highest classification accuracy of 92.15% is achieved by Resnet-50 model, the experimental findings highlight that the proposed CNN model outperformed the transfer learning approaches displaying accuracy of 97.5%. It can be concluded that the proposed CNN based multi-class grading method of OSCC could be used for diagnosis of patients with OSCC.

Introduction

Oral cancer, denoted as a category of head and neck cancer, includes major sub-regions of the lip covering oral cavity, nasopharynx, and pharynx (National Institutes of Health, 2018, WHO, 2017), consisting of about 85% of the category. Right off the bat, oral cancer is a life-threat disease due to the fact that its precursor symptoms and warning signs may not be observed by the patients habitually because of which this disease may rapidly progress into carcinoma stage within a short period.

The National Institute of Health, USA confirms oral cancer as the sixth most prevalent occurring cancer in the world (National Cancer Institute, 2018). The scenario is the same in India (16.1% of all cancers amongst men and 10.4% amongst women Sankaranarayanan et al., 2005) and alarming in the region of the study (44.5% for men and 16.7% for women Asthana, Patil, & Labani, 2016). While oral cancers are dominant in India compared to developed countries (Maxwell Parkin, Bray, Ferlay, & Pisani, 2001), buccal mucosa of the oral cavity has been reported as the most prevalent cancer site of the oral cavity in the same (Singhania, Jayade, Anehosur, Gopalkrishnan, & Kumar, 2015) next to tongue and gingiva which is subjective to 13,500 cases being reported by cancer registries in India against 1272 registered cases worldwide during the period 1990–1996 (Centre, Disease, Cancer, & Programme, 2013). In our case also, 80% of study cases primarily account to the buccal carcinoma group. This may be due to the high consumption of tobacco and betel nut. The estimate gives a reason to believe that the incidence rate related to buccal mucosa is increasing in upcoming years which is why buccal carcinoma is highlighted as the ‘site-specific’ oral cancers amongst the Indian population. Fig. 1 (Squirrel, 2012) shows the anatomical structure of the oral cavity as a whole.

Oral cavity cancers are also known to have a high recurrence rate compared to other cancers. Therefore, an in-depth exploration of either its staging or its grading is necessary for its prognostic treatment. More than 90% of cancers that occur in the oral cavity are squamous cell carcinomas (SCC). This cancer group is characterized by epithelial squamous tissue differentiation and aggressive tumor growth disrupting the basement membrane of the inner cheek region. Commonly, clinical procedures for prognosis and treatment are evaluated on Tumor-Node-Metastasis (TNM) staging. However, a five-year survival report based on oral cancer reveals a prognosis rate of approximately 35% to 50% (Iyer et al., 2004, Pradhan, 1989) ensuring quantitative histological grading of tumors, that incorporates the in-depth study of various pathological aspects related to SCC, as a more preferential method than tumor staging for increasing disease survival rate. Hence, from a pathologist’s point of view, providing precise histopathological identification in the context of multi-class grading is important. This gives a rationale to combat the issue by incorporating deep learning-based disease diagnosis or prediction methods with clinical prospective which are hot research topics nowadays (Lecun et al., 2015, Litjens et al., 2017). Oral SCC is morphologically categorized into Normal, Well-differentiated, Moderately differentiated and Poorly differentiated classes based on Broder’s system of histological grading (Doshi, Shah, Patel, & Jhabuawala, 2011). The cellular morphometry highlighting the tumor growth displays a very minute histological difference separating the three classes which are very hard to capture by the human eye. It has remained elusive due to its highly similar histological features which even pathologists find difficult to classify. Even though most oral SCCs are moderately differentiated, they all have different metastasize characteristics and implicate different prognosis, recurrence rate and survival, and treatment management. Akhter, Hossain, Rahman, and Molla (2011) and Fortin et al. (2001). Therefore, with the growth of healthcare standards all over the world, it is necessary for an overhaul of pathology, which would involve more rapid and accurate diagnosis.

Computer-aided diagnosis systems involve three stages (1) tissue segmentation highlighting Region of Interest (RoI) (2) disease quantification through the identification of tissue malignancy and (3) differential diagnosis. In this work, we have attempted the second task, viz. to quantify OSCC according to the four different histopathological grading, including normal cases. The conventional practice followed by pathologists is based on subjective assessment. Apart from that, there are many other factors which affect manual judgment. These may be the make/model and status of the microscope, the lighting conditions of observation or data acquisition, the quality of the stains/slides, experience of the pathologist, time devoted for every observation and so on. All such factors may lead to diagnosis error or delay, affecting follow up action (Betta et al., 1997). Whereas automated systems applying traditional machine learning entail analysis of all features, deep learning eliminates the need for domain expertise and hardcore feature engineering. With the availability of high-performance hardware like GPU, deep learning algorithms can also yield high accuracy for classification without manual feature representation of the input data that most machine learning algorithms do. In terms of treatment management, a biopsy image is considered a pathological gold standard. There are many articles in the literature related to deep learning using biopsy images that suggest the adoption of pre-trained models for classification task (Kassani et al., 2019, Srivastava et al., 2019, Xie et al., 2019). Therefore, it is expected that deep learning methods will work efficiently without selective feature engineering that conventional machine learning methods rely on discerning the minute and fuzzy difference between the sub-classes of OSCC.

Literature reveals few researchers have applied deep learning and machine learning techniques in the retrospective study for early diagnosis and survival prediction rates of oral cancer patients (Karadaghy et al., 2019, Kim et al., 2019, Mohd et al., 2015). But the majority of works doing prospective study highlights machine learning approaches for early detection of oral sub-mucous fibrosis (OSF) rather than OSCC cases (Chodorowski et al., 2000a, Krishnan et al., 2011, Krishnan et al., 2010, Patra et al., 2012, Prabhakar and Rajaguru, 2017). This is because OSF is considered to be chronic as the majority of cases transformed into oral cancer on progression. Studies on OSCC done till date have been on binary classification by Rahman, Mahanta, Chakraborty, Das, and Sarma (2018) and Rahman, Mahanta, Das, and Sarma (2020) using shape, color and texture features from biopsy images. Classification accuracy of 100% has been reported using SVM and Linear Discriminant classifiers.

A detailed literature study reveals that no work has been done so far on automated multi-class grading. As per our knowledge, this is the first work undertaking a comprehensive study on various deep learning architectures for the classification of multiple cells (or multi-classification) in oral carcinoma using OSCC biopsy images.

Based on prior reviews, the following underlying questions have been addressed which we feel is necessary for this study:

  • 1.

    Can existing deep learning models be fine-tuned with OSCC images using a transfer learning approach for classifying the images into its multiple classes?

  • 2.

    Can a deep learning model, built from scratch, be extended by learning more visual and hierarchical features for classifying the images into its multiple classes?

Section snippets

Materials and methods

Deep learning has been an active research domain in medical image analysis and applications with the successful winning of image recognition challenges (like ILSVRC grand challenge) (Esteva et al., 2017, Litjens et al., 2016, Roth et al., 2016, Shin et al., 2016). A classical CNN architecture consists of multiple convolutional layers, pooling layers and fully connective layers along with input and output layer. Subsequently, a series of architectures have come up following Lenet-5 (1998)

Performance metrics for evaluation of the classification task

The prediction made by a standard classifier is verified using evaluation metrics on multi-class classification (Sokolova & Lapalme, 2009) as shown in Table 4. Here, true-positives and true-negatives denote the number of positive and negative classes correctly predicted by a classifier; false-positive and false-negative denotes the number of positive and negative classes that are incorrectly predicted by a classifier.

Results

In this section, we present the experimental results to justify the hyper-parameters found ideal for our proposed CNN model. We also demonstrate the effect of classification accuracy on the pre-trained CNN models using three different optimizers. Additionally, comparative evaluation is made between pre-trained models and proposed CNN based on learning curve results to highlight the superiority of the proposed CNN method.

Discussion

Analysis of features of the oral epithelial tissue is the gold standard of clinical evaluation for grading of oral carcinoma. This qualitative evaluation requires skilled pathologists to analyze every minuscule difference between a healthy cell and a malignant or cancer cell. This makes the entire process laborious and time-consuming which may contribute to slow prognosis and treatment. Automated detection of OSCC grades can reduce such drawbacks and can also overcome observer-biases. A major

Conclusion

There are many components in the complete protocol of diagnosis of cancer by pathology. Histologic grading is the most important component to learn as much as possible about the cancer cells. It is the key to plan the best treatment and make a decision on overall outcomes and goals. Automation is the need of the hour to ascertain a rapid and accurate diagnosis, thereby paving the way to accommodate more patients. Current work highlights, probably for the first time, the huge potential of

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgment

This work is supported logistically by the Institute of Advanced Study in Science and Technology, Assam, an autonomous R&D institute under the Department of Science and Technology, Govt. of India, New Delhi, India . Authors would like to acknowledge Tabassum Y. Rahman and Dr B Barooah Cancer Institute, Guwahati, Assam for data acquisition.

References (51)

  • CentreN. et al.

    National centre for disease informatics and research indian council of medical research consolidated report of hospital based cancer registries an assessment of the burden and care of cancer Patients. (September)

    (2013)
  • ChenC.H. et al.

    Handbook of pattern recognition and computer vision. Handbook of pattern recognition and computer vision

    (1999)
  • ChodorowskiA. et al.

    Oral lesion classification using true-color images

    Doktorsavhandlingar Vid Chalmers Tekniska Hogskola

    (2000)
  • ChodorowskiA. et al.

    Oral lesion classification using true-color images

    Doktorsavhandlingar Vid Chalmers Tekniska Hogskola

    (2000)
  • DoshiN. et al.

    Histological grading of oral Cancer: a comparison of different systems and their relation to lymph node metastasis

    National Journal of Community Medicine

    (2011)
  • EngstromL. et al.

    A rotation and a translation suffice: Fooling CNNs with simple transformations

  • EstevaA. et al.

    Dermatologist-level classification of skin cancer with deep neural networks

    Nature

    (2017)
  • FortinA. et al.

    Does histologic grade have a role in the management of head and neck cancers?

    Journal of Clinical Oncology

    (2001)
  • Gedraite, E. S., & Hadad, M. (2011). Investigation on the effect of a Gaussian Blur in image filtering and...
  • HeK. et al.

    Deep residual learning for image recognition

  • IyerS.G. et al.

    Surgical treatment outcomes of localized squamous carcinoma of buccal mucosa

    Head and Neck

    (2004)
  • JeyarajP.R. et al.

    Computer-assisted medical image classification for early diagnosis of oral cancer employing deep learning algorithm

    Journal of Cancer Research and Clinical Oncology

    (2019)
  • KaradaghyO.A. et al.

    Development and assessment of a machine learning model to help predict survival among patients with oral squamous cell carcinoma

    JAMA Otolaryngology - Head and Neck Surgery

    (2019)
  • KassaniS.H. et al.

    Classification of histopathological biopsy images using ensemble of deep learning networks

    (2019)
  • KimD.W. et al.

    Deep learning-based survival prediction of oral cancer patients

    Scientific Reports

    (2019)
  • Cited by (104)

    • Predicting nodal metastases in squamous cell carcinoma of the oral tongue using artificial intelligence

      2024, American Journal of Otolaryngology - Head and Neck Medicine and Surgery
    View all citing articles on Scopus
    View full text