NIMG-44. INTEGRATING AUTOMATED LESION SEGMENTATIONS FROM SINGLE-IMAGES INTO ROUTINE CLINICAL WORKFLOW FOR VOLUMETRIC RESPONSE ASSESSMENT
Abstract INTRODUCTION Volume calculations have not been adopted into glioma response assessment due to lengthy times for manual definition and unreliable measures provided by automated algorithms. Relatively new artificial intelligence approaches such as convolutional neural networks have significantly improved lesion segmentation with performance accuracies >90%. However, their adoption into routine practice remains limited due to poor generalizability and failure rates approaching 25% when incorporated into clinical workflow. The latter can be attributed to 1) the requirement of four different types of anatomic images (T2, T2-FLAIR, T1 pre- and post-contrast); 2) cumbersome preprocessing including alignment, reformatting, and skull removal; and 3) the lack of a well-integrated clinical deployment system. The goal of this study was to demonstrate how simple modifications to a robust network coupled with an integrated workflow can provide reliable measures of tumor volume for real-time use in the reading room. METHODS Leveraging NVIDIA’s Clara-Train software and a molecularly diverse dataset of 400 labeled images for training, we modified a top-performing ensembled 2D-U-Net to require a single image-volume input (T2-FLAIR or post-contrast T1 for the T2-hyperintense or contrast-enhancing lesions) and deployed the results in the clinic to provide quantitative volumetrics. Inference was performed on a mix of image orientations without any reformatting or skull-stripping. RESULTS Training on only 115 of our 400 datasets, we achieved Dice Coefficients of 90% and 81% overlap of our auto-segmented T2 and contrast-enhancing lesions with manual labels in our 25-patient validation cohort (11 enhancing), compared to 91% and 83% overlap with the original model that required four anatomic images to segment each lesion. Radiologists can view segmentations directly from PACS as contours or overlays and provide numerical feedback for model refinement. The workflow has been applied on 50 cases to date without any failures and can be easily shared for deployment on any clinical PACS.