چكيده به لاتين
Abstract
According to the report by the International Agency for Research on Cancer (IARC): In the Global Cancer Report 2020, cancer was identified as the primary or secondary leading cause of fatal injury (ages 30 to 69) in 134 out of 183 countries worldwide. Lung cancer remains the leading cause of mortality among both men and women. According to IARC, the incidence of cancer is projected to rise from 18.1 million to 29.5 million between 2018 and 2040. Accurate and precise diagnosis of breast cancer (BC) is crucial for early detection and improved survival outcomes. Diagnosing the disease is generally challenging; however, machine learning plays a significant role in identifying the disease in individuals, monitoring their health, and recommending preventive measures.
Classical machine learning methods, such as decision tree algorithms, random forests, support vector machines (SVM), and simple two-layer neural networks, are not very effective in classifying image data. Deep learning (DL), a subset of machine learning (ML), operates directly on image data, enabling the automatic definition (learning) of appropriate features without human intervention. Therefore, the main objective of this research is to propose an efficient and robust deep learning model for breast cancer detection and classification.
In this thesis, to improve the accuracy of classification methods using machine learning, we present techniques based on deep learning and convolutional neural networks (CNN) on ultrasound breast images of women from the Kaggle dataset, which many studies have utilized. After data extraction, the steps of the image classification project (disease detection) include preprocessing, feature extraction, modeling, and evaluation. For the three-class image classification stage, improved CNN, VGG19 network, ResNet-50, ResNet-152, MobileNet, Xception network, and a proposed enhancement model were used for feature extraction and classification.
Due to overfitting, an enhanced EfficientNet-B3 model was employed for improvement. EfficientNet introduces a novel method for scaling CNNs called compound scaling. The model, due to its improvements in structure compared to other methods, delivered highly satisfactory results, achieving the best accuracy at 96.52%, with an observed improvement of up to 30% and 15% compared to other models.
