Abstract
BACKGROUND: The Zurich Pituitary Score (ZPS) is an externally validated radiological grading scale to predict the likelihood of gross total resection (GTR) on coronal T1w magnetic resonance imaging of pituitary adenomas. The ZPS is based on the ratio of maximum tumor horizontal diameter and minimum intercarotid distance and on carotid artery encasement. While the interobserver agreement of the ZPS was relatively good, automated grading would be beneficial. METHODS: A nnU-Net algorithm was trained to segment the manually labeled tumor tissue and the cavernous segment of the internal carotid artery. Subsequently, maximum horizontal tumor diameter and minimum intercarotid distance were extracted. Last, a seed-growing algorithm checked for encasement of the carotid to determine the ZPS. RESULTS: 213 patients were included, of which 128 (60%) had non-functioning adenomas, 49 (23%) a growth-hormone secreting and 19 (9%) a prolactin producing tumor. Accordingly, ZPS gradings were I = 63 (30%), II = 94 (44%), III = 41 (19%) and IV = 15 (7%). Dice score (mean ± standard deviation) for the tumor, left carotid, and right carotid in training validation of 0.78 ± 0.24, 0.62 ± 0.31, 0.62 ± 0.30 and during holdout testing of 0.79 ± 0.24, 0.59 ± 0.32, 0.58 ± 0.33 was reached. After the exclusion of two cases with poor segmentation results, intraclass correlation coefficients [95% CI] for the intercarotid distance, maximum horizontal tumor diameter, and the ZPS ratio of the two measurements were 0.89 [0.80, 0.94], 0.91 [0.82, 0.96], 0.80 [0.66, 0.89] respectively. Cohen's weighted Kappa for the final ZPS grading was 0.79 [0.68, 0.90] and Spearman rank correlation was 0.83. CONCLUSIONS: We developed and internally validated a machine learning-based method for fully automated grading of the ZPS. Generally, robust segmentation performance was achieved. While ZPS grading generally worked well, human ratings remain superior in many situations. Especially for raters with low experience, our approach offers a solid and objective alternative.