Biopsy cost reduction for early diagnosis of breast cancer using hybrid deep learning techniques

Breast cancer has become one of the most important diseases all over the world. Early diagnosis of breast cancer is very important to reduce the mortality rate. Breast biopsy is one of the most commonly used methods to diagnose breast cancer. However, breast biopsies are sometimes performed even when the patient does not have breast cancer. This leads to various problems for patients such as anxiety, pain, and healthcare costs, etc. Therefore, we proposed a hybrid model to reduce the cost of breast biopsies for early detection of breast cancer using deep learning and machine learning methods. Although many methods have been applied to a range of breast cancer diagnoses using one data type (image or text), the combination of three data types (image, text, and survey) has not been used directly for breast cancer diagnosis in the literature, and the problem of breast cancer diagnosis still needs to be improved. The proposed model combines three types of data, namely image (mammography), text (radiology report), and survey (patient history, physical examination, etc.) of each patient, and generates amalignant or benign output to identify the type of breast cancer. Therefore, a hybrid system consisting of three independent parts is described in this thesis: In the first part, deep learning models such as pre-trained models (VGG16, AlexNet, and ResNet50) and a transformer model are used to classify mammography images. In the second part, machine learning models with different combinations are used to classify surveys. In the third part, similar to images, pre-trained models (BERT versions such as BERTMultilingual, BERTClinical and BERTTurkish) and transformers are used to classify radiology reports. The ensemble model takes the results of each part as input and produces an output to diagnose breast cancer. We also propose a risk-based hybrid neuro fuzzy rule-based system to calculate the risk of breast cancer. We show that using three types of data can reduce the cost of breast biopsies in breast cancer diagnosis. The hybrid model with the hospital dataset improves the precision value by up to 100%. Key Words: Breast cancer diagnosis, Transformers, VGG16, AlexNet, ResNet50, BERT