Convolutional neural network can help differentiate FDG PET images of brain tumor between glioblastoma and primary central nervous system lymphoma

1855 Objectives Differential diagnosis of rapidly-growing primary brain tumors between glioblastoma and primary central nervous lymphoma (PCNSL) is clinically important because different treatment strategies are needed; glioblastoma requires total resection followed by chemoradiotherapy whereas PCNSL should be treated with chemoradiotherapy without total resection. FDG PET is useful for the differential diagnosis since PCNSL often shows stronger FDG uptake than glioblastoma. Many reports focused on the importance of the highest FDG uptake value in the tumor, such as SUVmax or tumor-to-normal ratio, but the voxels in the whole tumor have not fully used. Deep learning is a group of sophisticated machine learning techniques that have been recently introduced to computer vision and image recognition sciences. Since glioblastoma and PCNSL usually have different shapes, intensities, and levels of heterogeneity on FDG PET, we hypothesized that the deep learning technique may have potential to distinguish these two entities. Methods In this retrospective study, we reviewed a total of 31 brain tumor patients (either glioblastoma or PCNSL) who underwent FDG PET with a Siemens EXACT HR+ scanner. Of 31, based on visual assessment, 25 patients (12 glioblastoma and 13 PCNSL) showed higher uptake in the tumor than that in the surrounding brain tissues. A total of 409 brain slices containing tumors were extracted from the patients, consisting of 242 and 167 slices for glioblastoma and PCNSL, respectively. The number of slices per patient was 16.4 +- 6 (range, 7 - 28). An experienced nuclear medicine physician defined region of interest on each slice to enclose the entire tumor. Convolutional neural network (CNN), which is one of the deep learning techniques and is known to be feasible to image analysis, was employed. Leave-one-patient-out cross validation was conducted to evaluate classification performance, where CNN was trained using 24 patients data while the remaining patient was used for validation and the process was repeated 25 times. The algorism gave each slice a probability that the slice belongs to glioblastoma. The slice was considered as glioblastoma in case of >50% probability. The patient was categorized as having glioblastoma when the averaged probability of all the tumoral slices was greater than 50%. As a comparison, maximum standardized uptake value (SUVmax) of each patient was measured as another classifier by thresholding method with ROC analysis. Results First, the entire slices without ROI masking were given to the CNN program. On slice-based analysis, 72% glioblastoma and 65% PCNSL slices were correctly categorized, resulting in overall accuracy of 69%. On patient-based analysis, 75% glioblastoma and 62% PCNSL patients were correctly categorized, resulting in overall accuracy of 68%. Second, the manual ROIs were used to erase the non-tumoral regions, and then given to the CNN program. On slice-based analysis, 83% glioblastoma and 71% PCNSL slices were correctly categorized, resulting in overall accuracy of 79%. On patient-based analysis, all glioblastoma and 77% PCNSL patients were correctly categorized, resulting in overall accuracy of 88%. The overall accuracy of SUVmax was 80%. Conclusions Convolutional neural network may have potential for differential diagnosis of brain FDG PET between glioblastoma and PCNSL. The ROI masking significantly improved the diagnostic accuracy compared to the non-masking process, possibly due to physiological uptake of FDG in the normal brain tissues, and thus CNN will be needed to be combined with an automated appropriate tumor segmentation technique.