NIMG-36. AUTOMATIC STRATIFICATION OF ENHANCING AND NON-ENHANCING GLIOMAS INTO GENETIC SUBTYPES USING DEEP NEURAL NETWORKS AND DIFFUSION-WEIGHTED IMAGING
| Autor: | Julia Cluceru, Paula Alcaide-Leon, Yannet Interian, Joanna J. Phillips, Janine M. Lupo, Valentina Pedoia, Tracy Luks, Devika Nair, Javier Villanueva-Meyer |
|---|---|
| Rok vydání: | 2020 |
| Předmět: | |
| Zdroj: | Neuro Oncol |
| ISSN: | 1523-5866 1522-8517 |
| Popis: | INTRODUCTION Current WHO guidelines emphasize classification of diffuse gliomas by genetic alterations into three subgroups: 1) IDH-wildtype; 2) IDH-mutant, 1p/19q-codeleted; and 3) IDH-mutant, 1p/19q-non-codeleted. Non-invasive genetic characterization can benefit patients with inoperable lesions or who are administered molecularly-targeted therapy before surgery. Prior studies that use anatomical images and convolutional neural networks (CNNs) to distinguish either IDH-mutant from IDH-wildtype tumors, or 1p/19q-codeleted from non-codeleted tumors have resulted in misclassification of nonenhancing IDH-wildtype and enhancing IDH-mutant tumors. This study investigated the benefit of a priori separation of enhancing from nonenhancing lesions and the inclusion of ADC maps from diffusion MRI to genetic subgroup classification. METHODS 3D T2-weighted, T2-FLAIR, and post-contrast T1-weighted images were acquired preoperatively from 254 patients with newly-diagnosed gliomas. IDH1R132H mutations[VJ1] [CJ2], 1p19q-codeletions, ATRX alterations, and p53 mutations were assessed from the resected tissue to determine subtype stratification: IDH-wildtype (n=95), IDH-mutant, 1p/19q-codeleted (n=62), and IDH-mutant, non-codeleted (n=97). 3-channel input images were constructed for each patient using T2-FLAIR, T1-post-contrast, and either T2-weighted or ADC images. Three VGG-16 CNNs pre-trained on ImageNet were re-trained for: 1) lesions without enhancement, 2) enhancing lesions, and 3) all lesions together[VJ3]. RESULTS A network trained on only enhancing lesions predicted the IDH-wildtype subtype with the highest class accuracy (ADC 94%, T2-weighted 100%) compared to using all lesions combined (ADC 90%, T2-weighted 90%). Models trained using non-enhancing lesions and ADC yielded the highest accuracy classifying 1p/19q-codeleted/non-codeleted subgroups (87%/90% for the non-enhancing network vs 83%/81% for combined network). CONCLUSIONS Our results support a strategy that first considers whether a lesion is enhancing when predicting molecular subgroup and includes ADC if the lesion is non-enhancing. Analysis is underway to test this model framework on independent TCIA data. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|