Deep Learning Solutions for Lung Cancer Characterization in Histopathological Images

• Autor(a) principal: João Moranguinho Bastardo Moura
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://hdl.handle.net/10216/132624
• topic: Engenharia electrotécnica, electrónica e informática; Electrical engineering, Electronic engineering, Information engineering

Cancer is one of the leading death causes in the world, specifically, lung cancer. According to theWorld Health Organization (WHO), at the end of 2020, around 2.2 million people were diagnosedwith lung cancer, and 1.8 million fatalities resulted from it. Correctly identifying it's presence in apatient and classifying it's sub-type and stage is fundamental for the adoption of appropriate targettherapies. One of the gold standards used to identify and classify cancer is the microscopic visual in-spection of histopathological imagesi.e.small tissue samples excised from a patient. Expertpathologists are responsible for this inspection, however, it requires a significant amount of timeand sometimes leads to non-consensual results . With the growth of computational power and data availability, modern Artificial Intelligencesolutions can be developed to automate and speed up this process. Deep Neural Networks us-ing histopathological images as an input currently embody the state-of-the-art in automated lungcancer diagnostic solutions, with Deep Convolutional Neural Networks achieving the most com-pelling acuracies in tissue type classification. One of the main reasons for such results is theincreasing availability of voluminous amounts of data, acquired through the efforts employed byextensive projects like The Cancer Genome Atlas. Nonetheless, histopathological images remain weakly labelled/annotated, as most commonpathologist annotations refer to the entirety of the image and not to individual regions of interestin the patient's tissue sample. Recent works have demonstrated Multiple Instance Learning as asuccessful approach in classification tasks entangled with this lack of annotation, by representingimages as a bag of instances where a single label is available for the whole bag. Thus, we propose a bag/embedding-level lung tissue type and sub-type classifier using a Con-volutional Neural Network in a Multiple Instance Learning approach, where the automated inspec-tion of lung histopathological images determines the presence of cancer, and it's possible sub-type,in a given patient. Furthermore, we employ a post-model interpretability algorithm to validate ourmodel's predictions and highlight the regions of interest for such predictions.