Alocal Directional Ternary Pattern Texture Descriptor For Mammographic Breast Cancer Classification

ABSTRACT

Breast cancer is a top killer illness for women globally, but early and effective screening

can increase their survival rate. Mammography is the tool used by a radiologist to screen

for breast cancer, however, a radiologist is susceptible to human observer variability, and

therefore, reading and interpretation of mammography test results depend on the expertise

of the radiologist administering the test. To improve the reading and interpretation

accuracy of the test, researchers’ developed computer-aided extraction descriptors that

extract discriminant features. These descriptors include the Local Binary Patterns (LBP),

the Local Ternary Patterns (LTP), and the Local Directional Patterns (LDP), however, they

have not yet yielded satisfactory results in differentiating breast cancer tumor types. The

LBP descriptor is inadequately dependable in capturing breast cancer discriminant features

because it is easily affected by noise. The LTP descriptor uses a fixed threshold value for

all images in a dataset, making it not invariant to pixel value transformation. It is also not

practically easy to select an optimum threshold value in real application domains. The LDP

descriptor relies on top k significant directional responses and ignores the remaining 8-k

directional responses. Disregarding the remaining directional responses reduces the

computation efficiency since each pixel in an image carries subtle information. Given the

limitations identified among the mentioned local texture descriptors, developing an

effective texture descriptor becomes a viable and challenging research problem. Therefore,

this study seeks to develop an improved local texture descriptor that considers all

directional responses and applies an adaptive threshold in encoding image gradient. The

new Local Directional Ternary Pattern (LDTP) texture descriptor calculates the absolute

difference between the value of the center pixel and the values of its local neighboring

pixels for a 3x3 image region. To get edge responses in eight directions, the absolute

differences are convolved with a kirsch mask, then the pixels are transformed into zeros

and ones using mini-max normalization. We then passed the normalized values through a

soft-max function to get the probability of an edge in a certain direction. Then, two

threshold values are calculated and used to split the probability space into three parts for -

1, 0, +1 bits to generate a ternary pattern. The resultant Local Directional Ternary Pattern

(LDTP) code is then split into a positive and negative LDTP code. Histograms of negative

and positive LDTP encoded images are fused to get texture features. We validated the

LDTP texture descriptor on the Mammographic Image Analysis Society (MIAS) breast

cancer dataset using Support Vector Machine (SVM) and Artificial Neural Network

(ANN) classifiers for normal/abnormal and benign/malignant classes. When the LDTP

texture descriptor was compared against LDP, LTP, and other existing texture descriptors,

it showed robustness and reliability in encoding an image gradient. The highest

classification accuracy was attained by the SVM classifier, with 97.32% and 93.93% for

normal/abnormal and benign/malignant classes, respectively.