Work

Nimg-69. Distinguishing Radiation Necrosis From Brain Tumor Recurrence on Routine Mri: a Preliminary Human-machine Reader Comparison Study

Year: 2016

Type: article

Abstract: Distinguishing radiation necrosis (RN) from tumor recurrence (RT) on routine MRI is a major challenge in neuro-oncology. To date, there is no imaging modality that can reliably distinguish RT from RN. In this study, we quantitatively compare the performance of a machine learning classifier in conjunction with radiomic features (computer-extracted textures) on routine MRI (T1w, T2w, FLAIR) with the visual diagnosis of two attending neuro-radiologists in distinguishing RT from RN. 3T-MRI studies with T1w, T2w, FLAIR protocols in N=55 patients (22 RN, 33 RT) obtained at least 9 months post-chemoradiation were analyzed. 135 radiomic features (i.e. Haralick, Laws energy, histogram of gradients) within lesions were extracted for each MRI protocol. A random forest classifier was trained on N=40, and tested on N=15 using the extracted features. Independently, two expert board-certified neuro-radiologists (with 2-, and 4-years of experience) who were blinded to the pathology reports (ground truth), read the same MRI scans to diagnose the presence of RN or RT. The readers assigned a score between 0.5 (unclear) and 1 (completely confident) to each study, based on the confidence in their diagnostic reads. Analysis from classifier and radiologists’ reads were compared to the pathologic findings. Area under receiver operating characteristic curve was found to be 0.79 for classifier, 0.75 for expert 1 and 0.58 for expert 2. Consensus with correct classification between the two experts was only seen in 22 of 55 studies (kappa coefficient =0.47). Our results suggest that radiomic features provide complementary diagnostic information to capture subtle differences between RT and RN on routine MRI that may not be visually appreciable to expert readers. Once validated on a larger cohort, radiomic analysis could serve as a decision support tool to enable timely and appropriate patient management in brain tumors.

Source: Neuro-Oncology

DOI: 10.1093/neuonc/now212.580