The Impact of Double Transfer Learning in VGG Architectures for Metastasis Breast Cancer Detection

• Danyllo Carlos Silva e Silva UEMA
• Omar Andres Carmona Cortes UEMA / IFMA
• João Otávio Beira Diniz IFMA

Resumo

According to the World Health Organization (WHO), 2.3 million women were diagnosed with breast cancer in 2020, causing almost 700.000 deaths worldwide. The first occurrences (in situ stage) usually have a good response if there is an early detection because the earliest form of tumor does not have sufficient potential to kill. However, the next stage has a low survival rate because the most threatening tumor characteristic is cell division spreading throughout the body, damaging lungs, livers, bones, and the brain. An alternative to deal with this problem is to enhance the velocity of the diagnosis and detection of breast cancer. Thus, machine learning algorithms have proven to be effective in this task. In this context, this work investigates how double transfer learning improves two Deep Learning architectures, VGG-16 and VGG-19, using histopathological images, i.e., employing BreakHIS as the first transfer learning and PatchCamelyon as the second. Results indicate that using double transfer learning, results for precision, specificity, recall, and F-Score improve to 99%, 83%, 89%, and 90%, respectively.