Computer assisted diagnosis of basal cell carcinoma using Z-transform features

https://doi.org/10.1016/j.jvcir.2016.06.014Get rights and content

Highlights

  • A method for computer assisted diagnosis of BCC is presented.

  • The method provides the pathologist a consultative opinion in evaluating specimens.

  • For the first time, Fourier features are combined to form Z-transform features.

  • Experiments show that the results are in agreement with the pathologist's opinion.

Abstract

Detection of basal cell carcinoma tumor is of great importance for decision making in the disease treatment procedure. Visual inspection of the histopathological slides for tumor detection is laborious, time consuming and prone to inter and intra observer variability. In this paper, we have proposed an automated method for discriminating basal cell carcinoma tumor from squamous cell carcinoma tumor in skin histopathological images using Z-transform features, which were not used previously in image classification tasks. For the first time, it is shown that how two or three Fourier transform features can be combined to form one Z-transform feature. Experiments have shown that the tumor classification results obtained by our method are in reasonable agreement with the gold standards provided by expert pathologists.

Introduction

Basal cell carcinoma (BCC) is the most common form of skin cancer and accounts for approximately 70% of non-melanoma skin cancers [1], [2]. This epithelial malignant tumor consists of cells which look like the basal layer of the epidermis with a rather chaotic arrangement [3]. The diagnostic histological features, common for all types of the tumor, are basal cells with a thin pale cytoplasm surrounding oval nuclei with a granulated chromatin pattern. The tumor originates in the epidermis and invades the dermis in the form of solid nodules or projections with irregular forms, creating various growth patterns. BCC has a low metastatic potential, but if misdiagnosed or left uncured, can cause significant destruction and disfigurement by invading surrounding tissues. In some cases, the clinical diagnosis is not difficult for an experienced clinician. There are cases, however, that the final diagnosis may be determined only on the basis of a histopathological examination of a removed part of the tumor under the microscope.

Skin is composed of three primary layers: epidermis, dermis and hypodermis (subcutaneous adipose layer). Epidermis is divided into four layers: basal layer, squamous layer, granular layer and cornified layer. The two main layers, epidermis and dermis, are shown in Fig. 1. Basal cell carcinoma is caused by uncontrolled growth and proliferation of the cells in the basal layer.

Detection of basal cell tumor is not a trivial task, because of its similarity to some parts of squamous epithelium, as is shown in Fig. 2. Fig. 2a shows a skin histopathological image of BCC. The regions encompassed by red and green boxes are parts of basal cell tumor and squamous epithelium, respectively. Magnified views of these regions are shown in Fig. 2c and d. Moreover, tumor detection becomes a more challenging task when distinguishing the basal tumor versus squamous cell tumor is considered, an issue which is important in differential diagnosis of BCC against SCC. This is also shown in Fig. 2. Fig. 2b depicts a histopathological image of squamous cell carcinoma. The yellow1 box in Fig. 1a shows part of BCC tumor that has invaded deeper into the dermis and the yellow box in Fig. 2b shows a region which corresponds to SCC tumor, with magnified views shown in Fig. 2e and f. However, there are some fine differences between these two types of tumor: the former consists of cells which are more oval, darker and less mature (have a higher N:C ratio) with respect to the cells in the latter. In this paper, we address the problem of discriminating between these two types of tumor.

In the field of computer assisted diagnosis by processing histopathological images, the main challenge with the strategies which include nuclei segmentation is that all of the features used in diagnosis and grading stage depend on the accuracy of segmentation stage. On the other hand, due to the cluttered arrangement of the cells in skin histopathological images and presence of fragments in the cells membranes, especially in tumor regions, nucleus segmentation is an erroneous task. So it seems that the methods which are based on texture classification are more appropriate for the diagnostic process. In this paper, we have proposed such a method that learns the texture of the tumor by means of Z-transform features. This is the first work in which efficacy of using Z-transform in extracting features from images is explored.

The organization of this paper is as follows. A review of the works in the field of histopathological images analysis is presented in Section 2. The proposed technique is described in Section 3; results are demonstrated and discussed in Section 4, followed by conclusion in Section 5.

Section snippets

Related works

In the past decade, many authors have proposed automated systems for histopathological images processing to diagnose malignant diseases. Authors in [4], [5] proposed methods for cervical cancer detection and grading by color texture features and content based image retrieval techniques. The same problem is addressed in [6], [7] employing Gabor filters, marked watershed segmentation and Delaney triangulation. The method in [8] detects and grades oral cancer using higher order spectra features

Notes on auto-encoders

The auto-encoder feature extraction approach is based on the intuitively reasonable thought that features which best reconstruct the original images provide better discrimination power to classify images between different classes. In other words, the major visual content of the patterns is concentrated within these set of features. Particularly, considering the input image I of size h×w, an encoder function f:RdRn (with d=w×h) maps the image I to a n dimensional feature space and then, a

Experiment setup

For evaluation of the method, 33 histopathological images were obtained from skin tissue samples of unknown patients, 18 images of SCC (including SCC in situ) and 15 images of basal cell carcinoma. The tissues were stained using Hematoxylin and Eosin (H&E) and the images were digitally captured by Aperio ScanScope slide scanner from histopathology slides at 40× magnifications. From these high resolution images, regions of interest (regions which contain SCC or BCC tumor) are cropped by

Conclusion and future work

A method for detection of basal cell carcinoma in skin histopathological images using Z-transform features is proposed. It is shown that how two or three Fourier transform features are combined to form one Z-transform feature. The value of complex number pairs at which the Z-transform should be computed are determined from training image blocks. Evaluation results have shown that our method improves, to some extent, the results of tumor detection with respect to the previously proposed method.

Acknowledgement

The authors would like to thank Dr. Maryam Almassi, M.D., A.P.C.P, from Vanak Pathobiology Laboratory (Tehran, Iran) for insightful advices and helping us in medical aspects of this research.

References (42)

  • J. Arevalo et al.

    An unsupervised feature learning framework for basal cell carcinoma image analysis

    Artif. Intell. Med.

    (2015)
  • Alessandro De Stefano et al.

    Features of biopsy in diagnosis of metatypical basal cell carcinoma (Basosquamous Carcinoma) of head and neck

    Otolaryngologia Polska

    (2012)
  • David E. Elder et al.

    Lever’s Histopathology of the Skin

    (2007)
  • Y. Wang et al.

    Assisted diagnosis of cervical intraepithelial neoplasia (CIN)

    IEEE J. Sel. Top. Signal Process.

    (2009)
  • G. Rahmadwati Naghdy et al.

    Cervical cancer classification using gabor filters

    First IEEE Int. Conf. Healthcare Inform. Imag. Syst. Biol. (HISB)

    (2011)
  • G.H.B. Miranda et al.

    Structural analysis of histological images to aid diagnosis of cervical cancer

  • M.M. Dundar et al.

    Computerized classification of intraductal breast lesions using histopathological image

    IEEE Trans. Biomed. Eng.

    (2011)
  • M. Veta et al.

    Breast cancer histopathology image analysis: a review

    IEEE Trans. Biomed. Eng.

    (2014)
  • B. Akbar et al.

    Colon cancer detection based on structural and statistical pattern recognition

  • M.V. Dass et al.

    Classification of lung cancer subtypes by data mining technique

  • A. Tabesh et al.

    Multi feature prostate cancer diagnosis and Gleason grading of histological images

    IEEE Trans. Med. Imaging

    (2007)
  • Cited by (0)

    This paper has been recommended for acceptance by M.T. Sun.

    View full text