2026 Proffered Presentations
S238: ARTIFICIAL INTELLIGENCE APPLICATIONS TO DETECT CLIVAL REMODELING IN THE SETTING OF PITUITARY ADENOMAS
Diego D Luy, MD; Mousa Javidialsaadi; Andre Payman, MD; Faraz Behzadi, MD; Joseph F Zywiciel, MD; Andrew Pickles; Joseph Frazzetta, MD; Isaac Ng, MD; Shiau-Sing Cicierska; Vikram C Prabhu, MD; Chirag Patel, MD; Anand V Germanwala, MD; Loyola University Medical Center
Introduction: Pituitary adenomas may expand the sella and invade adjacent structures, including the sphenoid sinus and/or the clivus. Previously, identification of sellar remodeling assisted with detecting these tumors prior to the advent of computed tomography and magnetic resonance imaging. This study aims to quantify efficacy for discerning clival osseous changes in patients with pituitary adenomas when compared to controls using artificial intelligence and machine learning models.
Methods: Electronic medical records were retrospectively reviewed. Standard bone window CT head still images were analyzed using supervised machine learning/convolutional neural network (CNN) models trained on 3 axes (axial, coronal, or sagittal). CT sequences of a manually segmented clivus bone for each patient (102 images from 34 functioning pituitary adenomas, 240 images from 80 non-functioning pituitary adenomas, and 387 images from 129 normal patients)
Results: The model's most favorable performance was observed for axial sequences (Model 1: Axial Pituitary Adenoma vs Normal, accuracy 81%; Model 4: Axial Non-Functioning Pituitary Adenoma vs Functioning Pituitary Adenoma, accuracy 95%; Model 7: Axial Non-Functioning Pituitary Adenoma vs Functioning Pituitary Adenoma vs Normal, accuracy 78%). Added benefit of bilateral and anterior-posterior image features on axial views may have contributed to differences in performance. Although bilaterality of information was also available in coronal models, which consistently performed poorly compared to sagittal and axial sequences.
Discussion/Conclusion: No prior reports have documented use of CNNs to identify subtle osseous changes related to pituitary adenomas and potentially detect their presence based on CT bone windows alone. Our models yielded average accuracies of up to 81% with Model 1 (Axial Pituitary Adnenoma) vs control and 95% with model 4(Axial functioning vs nonfunctioning Pituitary Adenoma). Our models provide evidence that CNNs may be trained to provide acceptable levels of accuracy with CT imaging, a modality more readily available than MRI.
Figure 1: GRAD-CAM visualizations of CNN model emphasis in classification of groups, confusion matric, and ROC curve for Model 1 (Axial Pituitary Adenoma vs Normal)
Figure 2: GRAD-CAM visualizations of CNN model emphasis in classification of groups, confusion matric, and ROC curve for Model 2 (Coronal Pituitary Adenoma vs Normal)
Figure 3: GRAD-CAM visualizations of CNN model emphasis in classification of groups, confusion matric, and ROC curve for Model 3 (Sagittal Pituitary Adenoma vs Normal)
Figure 4: GRAD-CAM visualizations of CNN model emphasis in classification of groups, confusion matric, and ROC curve for Model 4 (Axial Non-Functioning Pituitary Adenoma vs Functioning Pituitary Adenoma)
