60103513

Computer vision for biomedical images

Sören Lukassen

Kommentar

Imaging techniques have become an integral component of both biomedical research and clinical practice. At the same time, automated image analysis has experienced significant progress, driving technologies ranging from image search to autonomous vehicles. This automation is increasingly applied to biomedical images. However, adapting computer vision algorithms for biomedical images introduces unique challenges due to their distinct properties uncommon in other imaging datasets.

In this course, you will explore the prevalent imaging modalities within biomedicine, including CT scans, MRI, ultrasound, and whole-slide microscopy images. We will investigate the common data formats for these images, understanding what distinguishes them from typical jpeg or png files, and how these distinctions can facilitate their efficient analysis. After an introduction to basic computer vision algorithms, we will trace the evolution of classification and segmentation models in this domain over the last decade, starting with convolutional neural networks and culminating with cutting-edge architectures such as vision transformers. Through practical exercises, you will apply your knowledge to a dataset of histology slide images from cancer patients, aiming to predict the tumor's stage and location. Additionally, we will investigate how your models arrive at their predictions, identifying the data patterns they consider informative and connecting these insights to the pathophysiological alterations within tumor tissues.

Our models will be built using the pytorch package in Python. While familiarity with coding neural networks is not a prerequisite, prior experience in general Python programming is expected.

By the conclusion of this course, you will be equipped to develop segmentation and classification models for biomedical images, recognize common pitfalls and artifacts, explain how your models arrive at their predictions, and effectively communicate your findings.

Schließen