This study presents a deep learning-based approach for accurate segmentation of pheochromocytoma (PCC), a neuroendocrine tumor, in abdominal CT scans. Instead of the conventional whole-body region-based annotation, various multi-class annotation strategies (11 in total) considering peripheral organs of PCC, such as liver, spleen, kidney, aorta, adrenal gland, and pancreas, were evaluated using the nnU-Net framework. Using the NIH Clinical Center dataset of 105 contrast-enhanced CT scans of 91 patients, the Dice Similarity Coefficient (DSC), Normalized Surface Distance (NSD), and instance-wise F1 score were used as evaluation metrics. As a result, the TKA annotation strategy including tumor, kidney, and aorta showed significantly higher performance than the conventional TB annotation strategy including tumor and whole-body regions in terms of DSC, NSD, and F1 score (p < 0.05). In particular, the F1 score was improved by 25.84%, and the accuracy of tumor volume measurement (R^2 = 0.968) was also superior. The superiority of TKA was also consistently demonstrated in the 5-fold cross-validation results. This study demonstrates that incorporating relevant anatomical information into the deep learning model improves the accuracy of PCC segmentation and aids in clinical evaluation and follow-up.
