Automatic Brain Tumor Image Analysis: Current status and opportunities
Mauricio Reyes, Raphael Meier, Simon Habegger, Nicole Porz, Philippe Schucht, Roland Wiest
With the advent of modern machine learning technologies and innovative technology transfer into the clinics, automated analysis of brain tumors from multimodal MR images has considerably improved. In this report we will present our contributions towards automated glioma segmentation from multi-sequence MRI that builds on the power of big data analysis and innovative concepts at the interface between clinics and biomedical engineering.
Through an evaluation on a database of 25 glioma patients, manually versus automatically generated tumor segmentations were compared for the complete tumor volume TV (enhancing, non-enhancing and necrotic tumor core), the TV+ (TV+edema), and the contrast enhancing tumor volume CETV. Significant differences for TV+ and TV (p0.05) were found, suggesting the value of the automatic segmentation approach to provide robust volumetric analysis of gliomas.
We will give special emphasis on lessons learned stemming from the design of these algorithms; their comparison through internationally renowned competitions; their dissemination worldwide, and clinical evaluations.
We will show how these advances can leverage longitudinal analysis of brain tumors, including analysis of pre-operative, post-operative, as well as follow-up multi-sequence MR images of brain tumor patients. To this end, a database of 6 cases presenting multiple time points (leading to a total of 36 volumes) were automatically segmented using a dedicated spatio-temporal segmentation algorithm. In comparison to manually segmented results, the results show a good agreement between the expert and the automatic segmentation, which yields results in only 5 minutes. Moreover, a higher robustness in comparison to using a single-time point 3D segmentation was observed. These results suggest the potential of using the available temporal information for precise and robust longitudinal segmentation of gliomas.
Finally, we will present our recent findings regarding volumetric analysis of brain tumors used as biomarkers for patient survivability analyses, where we present a new finding related to the high predictive power of the non-enhancing tumor component to patient survival. We will briefly discuss about the opportunities these new technologies can open for neurosurgery, radiomics, and others areas.Format: Poster