MRI based Deep Learning Networks for classifying 1p19Q co-deletion status in Brain Gliomas: A comparative study

Background

1p19Q co-deletion status is a crucial biomarker in brain tumor biology, significantly impacting therapy and prognosis. Determining 1p19Q status typically involves obtaining brain tissue from invasive procedures. We developed fully automated deep-learning networks for non-invasive classification 1p19Q status using both multi-contrast and T2-only MRI alone.

Methods

Multi-contrast brain tumor MRI were obtained from four publicly available databases (TCIA, LGG, UCSF & EGD) and three in-house/collaborator institutions (UTSW, NYU, UWM). Subjects were selected on the availability of multi-contrast or T2w MRI and ground truth 1p19Q status. Two separate 2D U-Nets (MCon-net & OT2-net) were developed using the nnUNet package. Subjects from TCIA+UTSW+LGG were used for training on a 5-fold Cross-Validation (CV) scheme with an 80/20 training/validation split. Trained networks were evaluated on true held-out test cases (NYU+UWM+UCSF+EGD) post CV.



Results

MCon-net achieved an accuracy of 89.5%, sensitivity of 62.4%, and specificity of 92.8% on the test data. Despite using only T2w MRI, OT2-net demonstrated comparable performance with an accuracy of 87.7%, sensitivity of 65.3% and specificity of 90.4%.



Conclusions

We demonstrate a good 1p19Q classification accuracy using both Mcon-net and OT2-net. This represents an important step towards non-invasive classification of 1p19Q status and clinical translation.