Purpose: To explore semi-automatic approach for segmenting regions of interest based on pixel level information.
Methods: Due to recent advances in machine learning and big data technologies, in radiotherapy field there is great interest in utilizing these advances to study treatment response, automating the organ segmentation, tracking moving tumors, online daily adaption of treatment plans and more. But in all of these studies possibility of achieving reliable results is directly proportional to reduction in inter-observer variability. We have developed a hybrid auto segmentation tool. Unlike atlas, model and deep learning based auto-contouring approaches users can interact in the automation process if necessary. To make it understand better our tool is like having a driverless car but rider can take control whenever it is necessary. Our INTeractive CONTouring (Intcont) tool draws contour around an edged region of interest (ROI). It combines automatic edge finding with the operator oversight and occasional intervention if the contour line deviates from the expected path. This is a universal tool and can be applied to segment tumor volumes as well as organs at risk. Moreover, Intcont allows to change local threshold at the edge if expert wants to include or exclude the actual edge from ROI (tighter or looser boundary around ROI).
Results: Intcont has successfully drawn reliable ROIs on all datasets we considered during development of the tool.
Conclusion: Intcont offers a novel approach to image contouring that provides combination of automated contouring with user interaction. It avoids contouring artifacts often introduced by fully automated contouring system and at the same time it is less subjective than outright manual drawing or manual editing when contours created by automated methods have artifacts. Rigorous analysis of the "correctness" of contours from Intcont will be done soon.