Deep learning vs. conventional methods for automatic quantification of total tumor radioactivity in positron projection images of mouse xenograft tumors

Positron projection imaging (PPI) of tumor-bearing mice under certain circumstances can provide accurate in vivo estimates of total tumor radioactivity, an important pharmacokinetic measurement. However, the number of images generated in these studies is typically very large and many 2D tumor regions-of-interest (ROIs) must be manually defined to obtain accurate radioactivity estimates. In this study, we compared several methods that might allow automatic quantification of tumor radioactivity content. In total, 120 images (n = 81 mice) were acquired in pairs during two separate experiments. The first experimental batch was used for development, and the second as an independent testing cohort. Four methodologies were evaluated, including deep-learning (U-net), region-growing (Level-Set), and thresholding (Otsu, mean value). For all methodologies, preprocessing of the images included uptake normalization to fixed window. Tumor radioactivity is defined as total uptake within a tumor region minus a background estimate. Performance metrics were evaluated for both segmentation results (Sorenson-Dice Coefficient) and radioactivity calculation results (Bland-Altman). Using the test batch data, DICE score for U-net segmentation was 0.82, vs. 0.5-0.6 for the other three methods. Bland-Altman plots showed a mean difference of -0.26 for U-net based calculations vs. -0.5 to -0.8 for the other methods. The U-net approach had the highest accuracy in both segmentation and subsequent radioactivity calculation.