Representation learning for breast cancer lesion detection

• Main Author: Raimundo, João Nuno Centeno
• Publication Date: 2022
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10400.26/43020
• topic: Breast Cancer Detection; Magnetic Resonance Imaging; Computer Vision; Machine Learning; Deep Learning; Convolutional Neural Networks

Breast Cancer (BC) is the second type of cancer with a higher incidence in women, it is responsible for the death of hundreds of thousands of women every year. However, when detected in the early stages of the disease, treatment methods have proven to be very effective in increasing life expectancy and, in many cases, patients fully recover. Several medical image modalities, such as MG – Mammography (X-Rays), US - Ultrasound, CT - Computer Tomography, MRI - Magnetic Resonance Imaging, and Tomosynthesis have been explored to support radiologists/physicians in clinical decision-making work- flows for the detection and diagnosis of BC. MG is the imaging modality more used at the worldwide level, however, recent research results have demonstrated that breast MRI is more sensitive than mam- mography to find pathological lesions, and it is not limited/affected by breast density issues. Therefore, it is currently a trend to introduce MRI-based breast assessment into clinical workflows (screening and diagnosis), but when compared to MG the workload of radiologists/physicians increases, MRI assess- ment is a more time-consuming task, and its effectiveness is affected not only by the variety of morpho- logical characteristics of each specific tumor phenotype and its origin but also by human fatigue. Com- puter-Aided Detection (CADe) methods have been widely explored primarily in mammography screen- ing tasks, but it remains an unsolved problem in breast MRI settings. This work aims to explore and validate BC detection models using Machine (Deep) Learning algorithms. As the main contribution, we have developed and validated an innovative method that improves the “breast MRI preprocessing phase” to select the patient’s image slices and bounding boxes representing pathological lesions. With this, it is possible to build a more robust training dataset to feed the deep learning models, reducing the computation time and the dimension of the dataset, and more importantly, to identify with high accuracy the specific regions (bounding boxes) for each of the patient images, in which a possible pathological lesion (tumor) has been identified. In experimental settings using a fully annotated (released for public domain) dataset comprising a total of 922 MRI-based BC patient cases, we have achieved, as the most accurate trained model, an accuracy rate of 97.83%, and subsequently, applying a ten-fold cross-validation method, a mean accuracy on the trained models of 94.46% and an associated standard deviation of 2.43%.