Pallavi Tiwari
A significant dilemma in the management of patients with brain tumors is the differentiation of a suspicious lesion on a follow-up MRI scan as tumor recurrence or radiation necrosis.1 These 2 conditions mimic each other clinically and radiographically. Even for highly experienced radiologists, the accurate identification of suspicious lesions on posttreatment MRI scans as either tumor recurrence or radiation necrosis is rather challenging. The lack of reliable, noninvasive techniques for the characterization of posttreatment lesions has led to the need to perform surgical interventions or invasive stereotactic brain biopsies for disease confirmation, which subjects patients to morbid side effects and mortality in rare cases.
Unlike artificial intelligence (AI) applications in breast and lung cancer diagnosis, the problem of posttreatment assessment has not been addressed by the AI community, potentially because of the underlying difficulty in distinguishing similarly appearing pathologies on imaging. This caught our attention, as our group’s research focus is on developing novel AI, machine-learning, and radiomic (subvisual mining of radiographic images) techniques for challenging clinical applications in oncology and neuroimaging. Following discussions with clinical collaborators, we became interested in seeing if AI and radiomic analyses could add value to expert interpretation. Our study included human-machine comparison on a limited cohort of holdout studies in which we demonstrated that while experts individually had an accuracy of 50%, the radiomics classifier was 80% accurate in distinguishing radiation necrosis from tumor recurrence. In follow-up studies presented at the 2016 RSNA2 and Society of Neurooncology (SNO)3 meetings, we demonstrated that our AI radiomics solution, in conjunction with experts, improved overall accuracy in distinguishing radiation necrosis from tumor recurrence to 92%.
Our study generated significant buzz in the neuroradiology community, with commentaries published in AJNR in October 2016.4 It was largely acknowledged that machine-learning methods (including the ones presented in our paper) are meant to “augment and not automate” clinical decision-making. This was also reflected in our results reported at the RSNA and SNO meetings, where we demonstrated the value that AI classifiers add in improving clinical reads when used in conjunction with expert reads. Our work subsequently received commendation from the Ohio General Assembly for making a scientific contribution toward improving brain tumor treatment management.
Since the publication of our research in AJNR, our group has patented a new radiomics technique, Co-occurrence of Local Anisotropic Gradient Orientations (CoLlAGe), to capture lesion heterogeneity.5 CoLlAGe computes entropy (measure of disorder) in pixel-level edge directions within the lesion.