Comparative Analysis of Deep Learning Models for Segmentation and Detection of Breast Lesions

• Autor(a) principal: Manuel Ferrete Ribeiro, Raúl
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/11110/2602
• topic: Breast Cancer; Deep Learning; Segmentation; Object Detection; Mammography

Breast cancer is the most common and deadly cancer in women worldwide. Early detection is crucial for the disease to be treated properly and with a lower mortality rate. Currently, mammography is the most effective type of imaging for detecting breast cancer at an early stage. Regular mammographic examinations, according to studies, can reduce mortality rates by 30-70% when breast cancers are diagnosed early, before they spread to other organs and tissues. However, most image analysis processes are still manual, and the outcome is potentially variable between observers. Automatic tumor detection and segmentation in mammographic images can aid in the diagnosis of breast cancer, making the process simple, fast and independent of the operator's experience. However, reliable detection and segmentation of lesions with this imaging modality is difficult due to the following factors: i) low contrast at lesion boundaries; ii) extremely varied lesion sizes and shapes; and iii) some extremely small lesions in the mammography image. Deep learning methods, particularly convolutional neural networks, have recently demonstrated high performance in different field of image processing. However, more research and development is needed before these techniques can be used with confidence in clinical practice. The main goal of this dissertation is to study deep learning methods to create a breast tumor detection and segmentation system for mammographic images. In order to achieve this goal, several segmentation and detection models present in the state of the art will be compared against a public CBIS-DDSM database. As a result, the best segmentation and tumor detection model is projected. Based on our comparison, we find that the value results are the MDA-NET and CenterNet based strategies with Dice of 93% for segmentation and a mAP of 70% for detection. Further, the tests performed on the segmentation and detection models produce results that surpass those seen in the literature, demonstrating the interest of these intelligent techniques.