In the radiotherapy treatment chain, segmentation of the Gross Tumor Volume (GTV) is one of the most crucial but also challenging steps. The challenge of manual GTV segmentation is reflected in inter-observer variation (IOV), where differences of millimeters up to centimeters can be observed, depending on the target site. Over the years, many efforts have been put into reducing IOV, for example by development of consensus guidelines, (inter-)national training, and the use of multi-modal imaging. Most often, a combination of each of these examples is needed, as introduction of a new modality in GTV segmentation can provide more data and insight, but also poses a risk for variation in interpretation of the new data.
Magnetic resonance imaging (MRI) is one of the obvious imaging tools to improve GTV segmentation, as it is known for improved soft-tissue contrast over CT imaging, but it can also provide more biological parameters of the tissue at hand, for example with diffusion weighted imaging (DWI), dynamic contrast enhanced imaging, and MR spectroscopy imaging.
By adding multi-parametric MRI to the imaging data for GTV segmentation, data interpretation becomes more complicated for human observers. Use of deep-learning to present the radiation oncologists with an automated GTV segmentation is therefore a logical next step. The challenge is in how to train and evaluate performance of these automatic segmentation tools. In the setting where we know CT imaging in combination with anatomical T1- and T2-weighed imaging leads to a reduction in IOV, training of the convolutional neural network (CNN) can be done in a supervised setting, using the clinical manual contours as ground truth (GT). However, when a novel MR-sequence is added, for example DWI, which has not been used in delineating the clinical GTV, discrepancies might occur of which can not directly be determined as a segmentation error of the automated tool.
This presentation will show an overview of current literature on using multi-parametric MRI in deep-learning segmentation of the GTV, including prostate, rectum, head and neck, and brain tumors. Furthermore, challenges in assessing segmentation performance will be discussed. Lastly, potential solutions for mitigating the challenges will be presented, including interactive deep-learning and pathology validation.