Human-centric ai for contouring in head-and-neck cancer therapy

Principal Investigators: dr. ir. Marius Staring (LUMC), dr. Anna Vilanova (TU Delft), dr. René van Egmond (TU Delft)

PhD students: Nicolas Chavez (TU Delft), Prerak Mody (LUMC)

Funding: HollandPTC-Varian 2019

Description

The aim of this project is to develop human-centric deep learning (artificial intelligence) methods to quickly and accurately locate and segment the tumor and organs-at-risk (OARs) based on planning as well as daily imaging. Specifically, we will focus on segmentation of the tumor and OARs in head-and-neck cancer. We envision a workflow where initially a fully automatic convolutional neural network predicts a first estimate of the segmentations. The result is presented to a human operator in such a way that human attention is directed to areas where input is most needed and/or most beneficial. Upon inspection, the human will give very limited feedback (cues), and never actually draw a complete contour, which is too time-consuming for IGOAPT. One may think of manipulation by fingers on a touch-screen, the use of scribbles, etc. Then, a second neural network will consider the first estimate together with the cues provided by the human, to produce an improved segmentation. By repeating this process a few times until the radiation oncologist is satisfied, a segmentation is obtained that can be approved for use in radiation planning. The combined workflow will be optimized, and is expected to drastically improve delineation and QA time in IGOAPT of head-and-neck cancer. Specific aims are:

To develop machine learning methods taking into account efficient human cues.
To develop efficient interaction and visualization mechanisms.
To optimize the human workflow protocol for use of AI in radiotherapy.
To develop software integrating these, and perform a user-study to evaluate their added value for head-and-neck proton therapy.