P04.15.B SEMI-AUTOMATIC INTERACTIVE RESPONSE ASSESSMENT FOR GLIOMAS WITH RANO2.0-ASSIST

Abstract BACKGROUND Recently, Response Assessment in Neuro-Oncology (RANO) criteria for high- and low-grade gliomas in adults were revised (RANO 2.0). Measurements can be time-consuming and subjective, resulting in increased inter-rater variability and reduced reproducibility. Automatic methods have been introduced in the past, however, existing tools are not based on the updated criteria, not freely available, and do not allow for interactive correction and validation of automatic results. To address these shortcomings, we developed a new tool called RANO2.0-assist (an extension of the research software 3D Slicer). RANO2.0-assist allows the user to modify, add, and remove bi-dimensional measurements, specify measurable and non-measurable lesions, and provide additional clinical information. Here, we evaluate automatic two-dimensional measurements provided by RANO2.0-assist as well as the response assessment classification by comparison with a manual ground-truth. MATERIALS AND METHODS A glioma dataset of 19 baseline MRI sessions with follow-ups was evaluated with RANO2.0-assist. Each session included contrast-enhanced T1-weighted, native T1-weighted, FLAIR, and T2-weighted scans. Manual two-dimensional measurements were provided by two expert radiologists, classifying the sessions into progressive disease (PD, n=13), stable disease (SD, n=6), partial response (PR, n=0), and complete response (CR, n=0). The RANO2.0-assist pipeline consists of: 1) image registration to a reference space 2) automatic segmentation of enhancing tumour tissue, oedematous tissue, necrotic tissue, and resection cavity 3) matching of lesions between baseline and follow-up 4) automatic measurement of bi-dimensional products and response classification according to RANO2.0 criteria. RESULTS On average, bi-dimensional measurements obtained with RANO2.0-assist were larger than manual measurements with an average relative difference of 21% ± 36%. This is because automatic bi-dimensional measurements capture narrow extensions of enhancing tissue that tend to be ignored in the manual ground-truth. However, this bias is partially canceled out in the computation of the response classification: Assuming the reference standard was correct, 15 cases (79%) were correctly classified without any alteration of the automated bi-demensional measurements, while 4 required alteration. CONCLUSION Our experiments show the potential for automatic response assessment but also highlight its pitfalls and need for human validation and correction of edge cases (human-in-the-loop). RANO2.0-assist provides an interactive user interface to facilitate this process. In the future, we aim to evaluate the impact on treatment response assessment through various options provided by RANO2.0-assist, such as post-processing of the automatic segmentations with morphological opening/closing operations.