Abstract
Artificial intelligence approaches to analyze pathological images (pathomic) for outcome prediction have not been sufficiently considered in the field of pituitary research. A total of 5,504 hematoxylin & eosin-stained pathology image tiles from 58 acromegalic patients with a good or poor outcome were integrated with other clinical and genetic information to train a low-rank fusion convolutional neural network (LFCNN). The model was externally validated in 1,536 patches from an external cohort. The primary outcome was the time to the first endocrine remission after stereotactic radiosurgery (SRS). The median time of initial endocrine remission was 43 months (interquartile range [IQR]: 13-60 months) after SRS, and the 24-month initial cumulative remission rate was 57.9% (IQR: 46.4-72.3%). The patient-wise accuracy of the LFCNN model in predicting the primary outcome was 92.9% in the internal test dataset, and the sensitivity and specificity were 87.5 and 100.0%, respectively. The LFCNN model was a strong predictor of initial cumulative remission in the training cohort (hazard ratio [HR] 9.58, 95% confidence interval [CI] 3.89-23.59; p < 0.001) and was higher than that of established prognostic markers. The predictive value of the LFCNN model was further validated in an external cohort (HR 9.06, 95% CI 1.14-72.25; p = 0.012). In this proof-of-concept study, clinically and genetically useful prognostic markers were integrated with digital images to predict endocrine outcomes after SRS in patients with active acromegaly. The model considerably outperformed established prognostic markers and can potentially be used by clinicians to improve decision-making regarding adjuvant treatment choices. © 2022 The Pathological Society of Great Britain and Ireland.