ESTRO 37 Abstract book

S9

ESTRO 37

concept has been carried forward in the on-going randomised PORTEC-4a trial that investigates molecular profile-based versus standard recommendations for adjuvant therapy, with the overall aim is to further decrease both over- and undertreatment in early stage endometrial cancer.

radiomics relies on even finer imaging details and automated processes to extract quantitative features from image sets using data-characterization algorithms, it is even more sensitive to the described challenges of MRI that have limited the successful adoption of MRI based dose painting. These challenges may be overcome but only through a concerted effort as a community to increase consistency and transparent communication around image acquisition, post-processing and analytic approaches. The standardization efforts of groups such as the Quantitative Imaging Biomarker Alliance (QIBA) and European Imaging Biomarkers Alliance (EIBALL) and the open science approaches of groups such as the Quantitative Imaging Network (QIN) are bringing us closer towards successfully utilizing quantitative imaging, including radiomics, for radiotherapy. SP-0028 Challenges of PET that have prevented adoption for dose painting and will persist for radiomics E. Troost 1 1 TU Dresden- Med. Faculty Carl Gustav Carus, Radiotherapy and Radiation Oncology, Dresden, Germany Abstract text Ever since the proposal of the biological target volume (BTV) by Cliff Ling in 2000 [4], the field of radiation oncology strived to integrate functional imaging of positron emission tomography (PET) into radiation treatment planning and selective dose enhancement. Today, 18 years down the line, PET has only found its way into target volume delineation of a few solid tumors, but has still not been adopted for radiation dose painting [2,10]. Here are some reasons for this. First, the most widely used PET-tracer, 18 Fluorodeoxyglucose (FDG) non-invasively depicts the tumor metabolism, but its use is hampered by uptake in normal tissues, such as brain and heart, and by tumour- or treatment-related (peritumoral) inflammation. Second, several hypoxia-related PET tracers, i.e. , 18F- misonidazole (MISO), 18 F-Fluoroazomycin arabinoside (FAZA), 18 F-HX4, have been developed and tested in clinical studies [9,15]. These differ regarding their specific kinetics, the optimal imaging and signal segmentation protocol, and most importantly, their availability is limited to a few dedicated centres [8]. With relatively long follow-up times to reach relevant endpoint (local control, overall survival), the value of repeat FMISO-PET imaging in head-and-neck squamous cell carcinomas was only shown recently in a cohort accrued between 2005 and 2013 [5,16]. Third, findings of PET tracers, e.g. , 18 F-fluorothymidine, may be influenced by the imaging protocols used and consequently show divergent results [1,3]. Finally, apart from these intrinsic differences, we have learned that volumes of increased tracer expression may vary over time, both without treatment and during radio(chemo)therapy, and that different tracer uptake volumes may not necessarily overlap [7,14,17]. Thus, there are numerous remaining questions, e.g. , when to best boost which volume using which tracer(s) and when to re-image to define the new boost volume. Alongside, initial enthousiasm of dose escalation was tempered by dose-limiting normal tissue toxicity, which now seems to be well-managed using moderately increased doses [6,13]. Furthermore, results of phase II dose-escalation studies in non-small cell lung cancer are awaited [11,12]. When moving away from “traditional” functional imaging to “radiomics”, the underlying characteristics of image acquisition and, more importantly, the tumor microenvironment will prevail. Thus, several of the challenges we have encountered will remain, and possibly new ones will appear. The presentation will give a brief

Symposium: Is radiomics going to deliver on the promises that dose painting made?

SP-0026 Machine learning in medical imaging: challenges and opportunities M. De Bruijne 1 1 ErasmusMC Rotterdam, Departments of Radiology- Nuclear Medicine- and Medical Informatics, Rucphen, The Netherlands Abstract text This lecture introduces current state-of-the-art techniques in machine learning in medical imaging and highlights the potential of these techniques for computer-aided diagnosis, quantitative image analysis, and therapy planning. We will discuss challenges to introduce these techniques in clinical practice, related to a requirement of annotated data for model training and to the difficulty of modeling variations in scan protocols. Examples of successful applications of machine learning in radiology will be shown. Abstract text There is great potential for MRI to guide and improve radiotherapy, well beyond the immediately notable superior soft tissue contrast. Used to its full potential, MRI has the ability to provide biological information that can characterize tumor and surrounding tissues based on cellular density, tissue perfusion and metabolism and beyond. These characteristics may reflect radioresistance within subregions of tumor and may be represented spatially through MR parametric maps. When these MR characteristics are gathered serially before, during and after treatment, changes in these MR parameters may serve as early response biomarkers. The dose painting approach hypothesizes that delivering differentially higher doses of radiation to radio-resistant regions, which may be better identified using multiparametric MRI, may improve tumor response and outcomes. Using information from serial MR images may facilitate adaptive dose painting to deliver lower doses to tumor subregions that are responding and higher doses to subregions that are not responding. A number of challenges have prevented wide-spread adoption of MRI based dose painting. These include variability in image acquisition across institutions and scanners, which can have significant impact on the consistency and quality of images used for planning. For instance, variability in the imaging protocol has the potential to impact the measured tumor volume, thereby impacting the radiotherapy target volumes and resulting radiation plan. Beyond image acquisition, the image post- processing and analysis can impact the results of quantitative MR parameters and the parametric maps used to help define the regions for dose painting. Finally, a major limitation in the use of MRI in dose painting is our limited understanding and knowledge around the pathology that is represented by the imaging data. As SP-0027 Challenges of MRI that have prevented adoption for dose painting and will persist for radiomics C. Chung 1 1 MD Anderson Cancer Center, Radiation Oncology, Houston, USA