Computer Outperforms Humans in MRI Brain Tumor Analysis

Computer analysis of subvisual data extracted from routine clinical magnetic resonance imaging (MRI) exams outperforms human experts at differentiating brain tumor recurrence from radiation necrosis, according to findings from a small study reported at the 21^st Annual Scientific Meeting of the Society for Neuro-Oncology, held November 17–20 in Scottsdale, Arizona.

Combining human and computer assessments might yield even better accuracy, reported Prateek Prasanna, a PhD student at Case Western Research University in Cleveland, Ohio.

“Radiomic features are independently diagnostic of tumor recurrence with an accuracy of 75%,” Prasanna said. “Radiomic analysis could serve as a decision-support tool to enable timely and appropriate patient management in brain tumors.”

Integrated radiomic analysis and expert diagnoses further increased recurrent tumor-detection and radiation necrosis detection.

Radiation necrosis is a delayed cancer radiotherapy injury to nontarget brain tissue that can emerge up to 6 months or longer after treatment. It can mimic tumor recurrence on standard MRI, with similar enhancement patterns.

“Definitive diagnosis is only possible via biopsy or resection,” Prasanna said. “There is a need for non-invasive techniques to differentiate recurrent tumors from radiation necrosis.”

The research team sought to evaluate the associations of histologic attributes of recurrent tumors and tumor necrosis with gradient orientation–based radiomic features extracted via high-throughput computing algorithms from routine clinical MRI image scan data. Statistical texture analyses were undertaken to quantify image characteristics like smoothness and heterogeneity. A goal of the study was to determine whether such computer-radiomic analysis can perform as well as human expert readers, and whether integrated radiomic plus expert consensus outperforms either alone.

Data from MRI T1 gadolinium, T2-weighted, and T2-FLAIR imaging exams were preprocessed and recurrent tumors and radiation-necrotic lesions were manually segmented. Radiomic subvisual features were extracted and their associations with pathologic features were analyzed.

Accuracy at differentiating necrosis from recurrent tumors (tumor detection accuracy) was 42% (5/12) and 50% (6/12) for two human readers, compared to 75% (9/12) for the radiomics classifier, Prasanna reported. The readers only agreed on 4 of 12 recurrent tumors.

Integrating expert reader diagnoses and radiomic classifier score improved detection accuracy for both recurrent tumors (91.7%; 11 of 12 cases) and radiation necrosis (100%; 3 of 3 cases).

The findings suggest that MRI-extracted subvisual features might reflect cellular differences between recurrent brain tumors and radiation necrosis. Prasanna and colleagues next intend to conduct prospective validation studies.