BREAST CANCER CLASSIFICATION USING MODIFIED CONVOLUTION NEUTRAL NETWORK TECHNIQUE

Abstract

Cancer is the abnormal growth of cells in the body tissue which can spread out rapidly in any part of the body. Some common types include lung cancer, colorectum cancer, prostate cancer and breast cancer, among which breast cancer is the most predominant. Breast cancer forms and grow uncontrollably in the surrounding tissue of the breast. The disease has resulted in several deaths most due to lack of early diagnosis and treatment. In breast cancer diagnosis, accurate detection and classification of the disease have been a major challenge over the years. A lot of time is wasted in making diagnoses due to huge volume of images to analyze as a result of numbers of increasing cancer cases. Some tissues have similar characteristics and formation often occurs in clusters varying between 0.05mm – 1mm in size making it difficult to locate, as a result, classification accuracy rate tends to decrease. Giant strides have been made in improving detection accuracy through application of machine learning algorithms like support vector machines, decision trees which have slightly improved system accuracy but limited when operating on raw image data. It requires features from the image to be first extracted before being fed into the model. Therefore, there is a need to design a system that can automatically learn features and make accurate prediction. The research focus on the use of a more recent approach in medical image analysis using a modified Convolutional Neutral Network (CNN) technique to learn, detect and classify the presence of breast cancer whether they are malignant (abnormal tissue) or benign (normal tissue) with a high precision. Transferring learning approach is adapted instead of building model from the scratch a method which has been proven to work satisfactory in breast cancer classification tasks. The techniques allow one to custom model tailored to a particular task using popular architecture as a baseline structure. The model uses the AlexNet architecture which was modified and tailored to our task. The work uses reflection and rotation as a form of augmentation technique to increase the amount of dataset that is used to train the model. The dataset undergoes a processing operation in order to enhance the low quality of the images before being fed into the model as training sources. The performance of the proposed model on test dataset is found to be; 95.80%, 95.00%, 80.00%, 92.30% and 93.63% for accuracy, sensitivity, specificity, precision and F1score respectively. The results show significant improvement in classification accuracy when compared with existing literature using deep learning techniques and MIAS breast cancer dataset. This work will help doctors in making accurate classification and reduced time wastage that is usually associated with the manual ways of analyzing breast cancer.