Clinical application of artificial intelligence algorithms in detecting clival remodeling in the setting of pituitary neuroendocrine tumors/pituitary adenomas.
Pituitary neoplasms may expand the sella and invade into adjacent structures, including the sphenoid sinus and/or the clivus. Previously, sellar remodeling assisted with detecting these tumors prior to the creation of computed tomography and magnetic resonance imaging. This project aims to quantify efficacy for discerning clival osseous changes in patients with pituitary neuroendocrine tumors (PitNETs) when compared to controls using artificial intelligence and machine learning models.
Electronic health records were reviewed. Still images of standard bone window CT heads were captured and compared using supervised machine learning/convolutional neural network (CNN) models trained on three singular axis (axial, coronal, or sagittal) CT sequences of a manually segmented clivus bone for each patient (102 images from 34 functioning PitNETs, 240 images from 80 nonfunctioning PitNET, and 387 images from 129 normal patients).
Overall, accuracies were favorable for axial sequences: Model 1 (axial PitNET vs. normal, accuracy 81%) and Model 4 (axial non-functioning PitNET vs. functioning PitNET, accuracy 95%), and Model 7 (axial non-functioning PitNET vs. functioning PitNET vs. normal, accuracy 83%). This performance difference may be due to added benefit of bilateral and anterior-posterior image features on axial views. Although this bilaterality of information is also available in coronal views, models consistently performed poorly compared to sagittal and axial sequences.
To date, no reports have detailed use of a CNN to identify subtle osseous changes and potentially detect PitNETs based on CT bone windows alone. Our models produced average accuracies up to 81% in correct identification of PitNET vs. control and 95% correct identification of functioning vs. nonfunctioning PitNET. These findings serve as a proof-of-concept that CNNs, may be trained to provide acceptable levels of accuracy with CT imaging, a modality more readily available than MRI.
Electronic health records were reviewed. Still images of standard bone window CT heads were captured and compared using supervised machine learning/convolutional neural network (CNN) models trained on three singular axis (axial, coronal, or sagittal) CT sequences of a manually segmented clivus bone for each patient (102 images from 34 functioning PitNETs, 240 images from 80 nonfunctioning PitNET, and 387 images from 129 normal patients).
Overall, accuracies were favorable for axial sequences: Model 1 (axial PitNET vs. normal, accuracy 81%) and Model 4 (axial non-functioning PitNET vs. functioning PitNET, accuracy 95%), and Model 7 (axial non-functioning PitNET vs. functioning PitNET vs. normal, accuracy 83%). This performance difference may be due to added benefit of bilateral and anterior-posterior image features on axial views. Although this bilaterality of information is also available in coronal views, models consistently performed poorly compared to sagittal and axial sequences.
To date, no reports have detailed use of a CNN to identify subtle osseous changes and potentially detect PitNETs based on CT bone windows alone. Our models produced average accuracies up to 81% in correct identification of PitNET vs. control and 95% correct identification of functioning vs. nonfunctioning PitNET. These findings serve as a proof-of-concept that CNNs, may be trained to provide acceptable levels of accuracy with CT imaging, a modality more readily available than MRI.
Authors
Luy Luy, Javidialsaadi Javidialsaadi, Payman Payman, Behzadi Behzadi, Zywiciel Zywiciel, Pickles Pickles, Frazzetta Frazzetta, Ng Ng, Cicierska Cicierska, Prabhu Prabhu, Patel Patel, Germanwala Germanwala
View on Pubmed