Introduction developing automatic reading procedures of such biopsies [2-4]. The appropriate automatic reading
1. Introduction developing automatic reading procedures of such biopsies [2-4]. The appropriate automatic reading
Colon cancer is one of the most frequent cancers technique must be carefully selected to process affecting men. It is characterized by abnormal and microscopic images that are high resolution, gray uncontrolled cellular proliferation. Detecting a cancer scale and multispectral. Following detection of cells at an early stage is of paramount interest since its within an image, the system must extract some treatment can be more guarantied than a metastasis characteristic parameters in order to distinguish cancer [1]. Many macroscopic and microscopic cancerous from normal cells [5, 6]. techniques may be used to detect colon cancer. The effectiveness of an automatic reading method is Medical imaging techniques such as MRI, CT scan, generally assessed by its capacity to analyze and PET and SPECT are considered as important tools to interpret a large number of images in a short time. detect, localize and estimate the volume of certain Different segmentation methods are available, tumors. Nevertheless, these tools are not able to detect including active contour or “Snake”. Although the last early cancer at the cellular level. Manual search of method is flexible and can be used as a non-gradient abnormalities in medical images is laborious and method that is important in the case of high resolution error-prone. Therefore, there is a need for an and heterogeneous images, its main disadvantages are automated computational technique to efficiently time consumption and detection of wrong contours quantify, distinguish and classify cellular images. during early iteration [7-9]. The authors have Thus, many recent works have been concerned by developed a new approach aiming to detect colon
cancer cells [10]. Our detection approach was derived Corresponding author: Jamal Charara, Ph.D., professor, research field: biomedical engineering. E-mail: [email protected].
from the “Snake” method but using a progressive
359 division of the dimensions of the image to achieve
Automated Classification of Segmented Cancerous Cells in Multispectral Images
and a circle having the same area and center of mass faster segmentation. The proposed method allowed
as of the cell. Fig. 1 shows the three steps required to accurate and efficient segmentation of images determine this parameter for the three cell types BH, containing distinct objects in a very short time [10].
IN and CA. The first row in this figure shows the On the other hand, the most important and yet most
segmented cells, while the second row shows the tedious tasks performed during microscopic analysis
applied circles. The area of the white region in the of cellular images is the classification of observed
third row illustrates the numerical value of the Xor cells into cancerous and non-cancerous. Automatic
cell-circle parameter.
cellular classification approaches include supervised The same methodology in Xor cell-circle is used to and non-supervised methods [11-13]. Supervised
estimate the Xor cell-convex (Fig. 2). The Xor classification can be distribution free, i.e. does not
operator here is between the cell and a convex which require knowledge of any priori probability covers the cell. distribution functions, or statistical techniques based
Similarly, the Xor cell-rectangle is an operator on probability distribution models that may be
between the segmented cell and a rectangle covering it. parametric (such as Gaussian distributions) or The SD of the positions of the contour points is: non-parametric. Nevertheless, probability distribution
(1) models cannot be considered as real models and may
SD 2
lead to wrong classifications. Non-supervised classification techniques attempt to identify clusters or
natural grouping in the feature space. The main disadvantage here is that two close distributions may
be identified as one class [14]. The objective of the present paper was to classify the different cell types based on several morphological parameters and on probabilistic neural network. Three types of cells were used to assess the efficiency of our
classification model, including Benign Hyperplasia (BH), Intraepithelial Neoplasia (IN) that is a precursor state for cancer, and Carcinoma (Ca) that corresponds to abnormal tissue proliferation (cancer).
(a) (b) (c)
2. Materials and Method
Fig. 1 The three steps required to determine the Xor cell-circle parameter for the three cell types: (a) BH; (b) IN;
2.1 Morphologic Parameters
and (c) CA. The first row shows the segmented cells, while the second row shows the applied circles. The area in the third row
Nine morphologic and characteristic parameters
illustrates the value of Xor-circle parameter.
were used to classify segmented cells, including area and perimeter of the cell, Xor cell-circle, Xor
cell-convex, Xor cell-rectangle, standard deviation (SD) of the positions of the contour points, deviation sum (DS), eccentricity (E) and solidity (S) of the
(a) (b) (c) detected cell.
Fig. 2 (a) Segmented cell; (b) convex covering the cell; (c) Xor
Xor cell-circle operator is applied between the cell
cell-convex.
Automated Classification of Segmented Cancerous Cells in Multispectral Images
where, N is the number of the contour points, X i is the
2.3 Algorithm
distance between a contour point i and the center of The developed Matlab algorithm includes two the cell and X is the mean value of X i . successive parts. First, it receives the microscopic The distance from each point of the contour to the sequential band image. Specification of the desired mean contour is summed to determine the DS as band of the image allows the algorithm to provide a
follows:
final segmented image. Second, the algorithm extracts DS
X X (2)
the morphological parameters of the segmented cell To measure E, an ellipse is interposed onto the cell
i 1
that are used as input for the probabilistic neural to cover it. The E parameter is given then by the
network to detect the type of cell. The best three shape following equation:
parameters of the cell classification were used as input
of the probabilistic neural network that was applied on where f 1 and f 2 are the two foci of the ellipse and L is
E = distance (f 1 ,f 2 )/L
four unknown cell types.
its major axis length. Fifty four images of biopsies, abnormal and known Finally, S is a scalar specifying the proportion of
cell types (BH, IN and Ca) as well as four images of the pixels in the convex hull that are also in the region
unknown cell types were used. Images were of the cell. It’s computed as follows:
segmented and observed on a midrange of 575 to Solidity = Area (cell) / Area (convex)
584.375 nm.
2.2 Neural Network Classification Methods
3. Results and Discussion
Several methods and algorithms allow classification Table 1 showed the values of the nine parameters using neural network [14]. The probabilistic neural
extracted from the three cell types of Fig. 1. It was network was used here, in which the distance
clear that among these parameters, only three of them, separating a test variable from known class variables
including area, Xor cell-convex and S, can better is measured and classified as to be from the nearest
separate the three cell types. Fig. 3 showed the class.
distribution of the three cell types using these three The activation function is used to introduce
parameters. Although the three types were well non-linearity in the functioning of the neuron. The
separated in three regions, some IN cells were close to activation function of the probabilistic neural network
BH cells.
is a function that measures the distance of unknown The parameters Area, Xor cell-convex and S were variable to all known class variables. This activation
used as input in the probabilistic neural network function plays the role of a low pass filter.
classification procedure and applied on four unknown
Table 1 Cell morphological parameters.
Parameter
BP cell
IN cell
Ca cell
Area (pixels)
166866 Perimeter (pixels)
5650.10 Xor cell-circle (pixels)
58254.87 Xor cell-convex (pixels)
58605.12 Xor cell-rectangle (pixels)
87465.62 Standard deviation (pixels)
Deviation sum (pixels)
263492.51 Eccentricity
Solidity
Automated Classification of Segmented Cancerous Cells in Multispectral Images
detection and classification of cancer cells based on our newly developed segmentation approach [10].