Glioblastoma segmentation using unsupervised machine learning

 

Ziel/Aim Multiparametric PET/MRI data of glioblastoma is complex and requires thorough understanding in order to produce: accurate tumor delineations, reliable segmentations of tumor heterogeneity, predict aggressiveness, stratify risk and therapy response assessments. Supervised machine learning (ML) approaches are not easily interpretable. Here, we focused on the segmentation using multiparametric MRI using an unsupervised ML approach.

Methodik/Methods 50 publicly available patient datasets belonging to the BRATS competition 2016 were used for analysis [1]. The datasets consist of T2 FLAIR, T2- and T1-weighted images pre- and post-injection of contrast agent. Pathological regions using provided radiological labels from the competition were provided per patient, which served as ground truth. We applied a gaussian mixture model with an increasing number of clusters until the calculated dice coefficient on the pathological region best approximated the ground truth.

Ergebnisse/Results Our results show a median detection accuracy of 0.996, a positive predictive value of 0.80, and a dice coefficient of 0.68 using 9 clusters to segment the data. The tumor regions are characterized by increments in signal intensity in pathological region, predominantly contrast enhanced and T2 FLAIR, however to a lesser degree than cerebrospinal fluid and necrotic regions. Healthy gray matter and white matter regions, with different grades of signal intensity were successfully discriminated from pathological regions.

Schlussfolgerungen/Conclusions Our results show multiparametric data using gaussian mixture models can produce accurate and specific tumor segmentations. However, this methodology still lacks accurate sensitivity for all pathological tissue, and though it can serve as a tool to understand the multidimensional variations in the tissue and their relationship, it still does not capture all of the tissue of interest. Following steps involve comparisons to supervised ML methods and addition of specific PET radiotracers.

• Literatur/References

• 1 Proceedings of MICCAI-BRATS. 2016 . https://www.cbica.upenn.edu/sbia/Spyridon.Bakas/MICCAI_BraTS/MICCAI_BraTS_2016_proceedings.pdf last accessed 03.11.2019.
  PubMed