ESTRO 36 Abstract Book

S313 ESTRO 36 _______________________________________________________________________________________________

shown that additional information is present that is not necessarily observable by the human eye. Even more, computer vision algorithms have already shown promise in predicting crucial clinical or biological characteristics based on medical images. I will present an update on the latest work in radiomics and radiogenomics for clinical outcome prediction. In addition, I will provide a gentle introduction to deep learning and its potential to rapidly influence quantitative imaging analysis and eventually treatment prediction. SP-0597 Tissue classification models for prostate based on imaging and non-imaging data U. Van der Heide 1 1 Netherlands Cancer Institute Antoni van Leeuwenhoek Hospital, Radiation Oncology, Amsterdam, The Netherlands Imaging data form the basis for target definition in radiotherapy. To attain the best possible understanding of the tissue, increasingly a combination of CT, PET and multiple MRI sequences are used. For dose differentiation treatments, such as dose painting, it is necessary to use the same images to characterize the heterogeneity within the target volume. With quantitative feature extraction, this process reaches a new level of sophistication. Intensity, shape and texture features provide a characterization of the images that can be used to build a classifier that characterizes the disease on a voxel-by- voxel basis. For prostate cancer, multi-parametric MRI is used routinely for detection of tumors inside the gland. Using feature extraction techniques, we constructed a classifier predicting the presence of cancer. This classifier can be used either to facilitate target delineation or to apply directly for dose painting by numbers. While the performance of image-based classifiers is quite good for the peripheral zone of the prostate gland, it can be challenging to classify tissue between cancer and non- cancerous voxels in the transition zone. Confounders, such as benign prostate hyperplasia, exhibit similar imaging features as cancer and are thus hard to distinguish. In clinical practice, a radiation oncologist has more information about the patient’s disease to be used to improve the quality of target delineation. The a-priori prevalence of the distribution of prostate cancer in the gland is well known. For example, prostate cancers mostly occur in the peripheral zone of the gland and less in the transition zone. As part of their diagnostic work-up, patients have received biopsies in the gland, proving cancer presence. The distribution of positive and negative biopsies of a particular patient is available for a radiation oncologist and is considered when defining a GTV delineation. In a study of two independent cohorts of patients, we show that a classifier that combines a-priori prevalence and biopsy data with features derived from multi-parametric MRI, performs significantly better than a classifier based on imaging data alone. Imaging features are not only used for tissue classification, but also to construct prognostic and predictive models for outcome after treatment. For prostate cancer, similar models are constructed based on T2-weighted MRI, or even multi-parametric MRI. Again, addition of non-imaging data may improve the performance of such models.

Fig. 1 Example of 3D Doppler image for a cervical cancer patient. Grey scale represents the morphology, while color scale is related to the presence of flow in the corresponding area. SP-0596 Machine learning and bioinformatics approaches to combine imaging with non-imaging data for outcome prediction O. Gevaert 1 1 Stanford University School of Medicine, Biomedical Informatics Research, Stanford, USA Radiomics and radiogenomics are burgeoning fields of science that put quantitative analysis of medical images (CT, MRI, etc.) central in the analysis towards the goal of precision medicine. The idea is to extract quantitative information from images that can be used for tailoring treatment decisions to the individual patient. More specifically, radiomics is defined as the quantitative analysis of medical images by semi-automatically, and increasingly more automatically, extracting image features from images . Radiogenomics (also known as imaging genomics) is concerned with the mapping of high dimensional molecular data (e.g. transcriptomics, genomics) with quantitative image features that result from radiomics pipelines. Recent developments in both areas of radiomics and radiogenomics are changing the paradigm of precision medicine. While previous work has focused mainly on molecular analysis of cancer, radiomics and radiogenomics propose to harness the power of quantitative medical imaging. This has several advantages, medical images are part of the diagnostic routine, are increasingly more available digitally, and are non-invasive. Especially the latter, the non-invasive characteristic, provides translational opportunities for diagnosis and also in vivo therapy follow-up. For example, radiomics signatures that predict prognosis (e.g. recurrence) can be more easily translated without incurring extra costs. Additionally, radiogenomics allows mapping molecular pathway activities to image signatures for non-invasive assessment of pathway activities, and subsequently hypothesis for targeted treatment. Within the field of machine learning, deep learning and convolution neural networks (CNN) have recently revolutionized analysis of images with many far-reaching applications outside of medicine. More recently, deep learning has entered the medical domain, especially in medical imaging. While most applications have focused on pathological/histological images, applications on medical images are emerging. Currently being used mostly by radiologists to interpret disease and quantitative analysis is still limited. In the meantime quantitative analysis as done in radiomics and radiogenomics research has already

Debate: Debate: Precision in radiotherapy: mission complete!

SP-0598 Precision in RT: mission completed! A. Duffton 1 , C. Dickie 2 1 Inst. of Cancer Sciences-Univ. Glasgow The Beatson West of Scotland Cancer Center, Research & Development Radiographer, Glasgow, United Kingdom