Pseudo progression identification of glioblastoma with dictionary learning

Abstract

Objective: Although the use of temozolomide in chemoradiotherapy is effective, the challenging clinical problem of pseudo progression has been raised in brain tumor treatment. This study aims to distinguish pseudo progression from true progression.

Materials and methods: Between 2000 and 2012, a total of 161 patients with glioblastoma multiforme (GBM) were treated with chemoradiotherapy at our hospital. Among the patients, 79 had their diffusion tensor imaging (DTI) data acquired at the earliest diagnosed date of pseudo progression or true progression, and 23 had both DTI data and genomic data. Clinical records of all patients were kept in good condition. Volumetric fractional anisotropy (FA) images obtained from the DTI data were decomposed into a sequence of sparse representations. Then, a feature selection algorithm was applied to extract the critical features from the feature matrix to reduce the size of the feature matrix and to improve the classification accuracy.

Results: The proposed approach was validated using the 79 samples with clinical DTI data. Satisfactory results were obtained under different experimental conditions. The area under the receiver operating characteristic (ROC) curve (AUC) was 0.87 for a given dictionary with 1024 atoms. For the subgroup of 23 samples, genomics data analysis was also performed. Results implied further perspective on pseudo progression classification.

Conclusions: The proposed method can determine pseudo progression and true progression with improved accuracy. Laboring segmentation is no longer necessary because this skillfully designed method is not sensitive to tumor location.

Keywords: Dictionary learning; Diffusion tensor imaging (DTI); Fractional anisotropy (FA); Genomics analysis; Glioblastoma multiforme; Pseudo progression.