Application of Artificial Neural Network and Texture Features for Follicle Detection

Abstract

Follicles are fluid-filled sac found in women reproductive system. Follicle detection in ultrasound images of the ovary is vital for fertility treatment. They are normally detected manually by Gynaecologists for disease diagnosis and to track follicular development. This process is usually hectic and prone to error. The existing automated methods for the detection of follicles are fraught with low detection rates due to the presence of image artifacts and noises resulting from blood vessels, endometrium, and tissues as captured by the ultrasound machine. These impacts negatively on the accuracy of the existing automated systems. Ultrasound images of the ovary exhibit different echo-texture patterns for different objects including follicles, artifacts, speckle noises and other tissues. This research employed Gray Level Co-occurrence Matrix (GLCM) technique to extract second order texture features for the various objects present in the image. Further, Multi-Layer Perceptron (MLP) was employed to classify the detected objects based on the extracted texture features into follicles and non-follicles. The developed algorithm yielded an accuracy of 96%, sensitivity of 99% and specificity of 93%. Also, Follicle Detection Rate (FDR), False Acceptance Rate (FAR) and False Rejection Rate (FRR) were computed to be 98.94%, 7.00% and 1.00% respectively.