06-19, 11:00–11:45 (Europe/London), Tower Suite 1
Recent progress in deep learning for medical imaging has led to impressive results. Among them is a fully automatic human organ segmentation from Computed Tomography (CT) scans. Organ segmentation can be the end goal in itself, e.g. when it is directly viewed by clinical teams. It can also serve as an input to diagnostic aid tools. Moreover, specific knowledge can be extracted out of segmentations to build databases. These databases can then be used for reasoning about the anatomy or planning treatment.
In this talk, we will describe a multi-stage pipeline for processing CT scans for abdominal aortic aneurysm (AAA) treatment planning. We will share our experience in sub-organ multilabel segmentation. We will discuss the challenges with common loss functions, and with metrics not being well aligned with clinical significance. We will show how enhanced segmentation can be used to represent patient anatomy in an accessible way for end-users who plan treatment for new patients.
One of the most impactful applications we can build with deep learning (DL) are tools for medical imaging analysis.
In this talk, you will learn about a DL-based tool for aortic care. Automated aorta segmentation removes lots of manual effort but this is just the beginning of what we can achieve with such data. The end goal is to understand which treatment options are possible and most likely to help an incoming patient. To do so we need to extract structured knowledge from the patient’s anatomy and the previous cases in our database. We will describe our clinical decision support system. It is a tool for creating a database of historical cases in a form that makes it easy to inform treatment decisions for new cases.
The vascular system is different from other areas of the human body. A lot of information is contained in the shape (morphology) of the blood vessels and not necessarily in the contents of the organ, which transmits blood. That’s why a segmentation and specifically its shape is very important to analyse.
First, we will cover the challenges in 3D medical image segmentation such as: - data preparation - solving technical challenges with large models operating on large 3D arrays of data - a conceptual mismatch between the perceived quality of a segmentation model and commonly used volume-based metrics like the Dice coefficient
We will then discuss: - techniques which incorporate shape prior constraints to the learning algorithm - using domain knowledge to drive data representation
We will wrap up with an example of a multi-stage pipeline that incorporates semantic segmentation of CT scans, intensity-based measures, geometric features, and domain knowledge. The pipeline aims to guide clinical planning by finding similar patients in our database.
No previous knowledge expected
Tomasz is a Machine Learning Engineer at Cydar Medical. He works on variety of medical imaging analysis - segmentation, classification models, combining all that to shape the future of the software supporting endovascular aortic repair. He contributes to all the stages of the machine learning project - from the problem statement definition, through literature review and experimentation, up to deployment & monitoring. In the previous years he was developing large-scale ranking models in an e-commerce platform. Before focusing on machine learning he worked in general software engineering.