OBJECTIVE: Pediatric adamantinomatous craniopharyngiomas (ACPs) are histologically benign brain tumors that confer significant neuroendocrine morbidity. Previous studies have demonstrated that injury to the hypothalamus is associated with worsened quality of life and a shorter lifespan. This insight helps many surgeons define the goals of surgery for patients with ACP. Puget and colleagues proposed a 3-tiered preoperative and postoperative grading system based on the degree of hypothalamic involvement identified on MRI. In a prospective cohort from their institution, the authors found that use of the system to guide operative goals was associated with decreased morbidity. To date, however, the Puget system has not been externally validated. Here, the authors present an interrater reliability study that assesses the generalizability of this system for surgeons planning initial operative intervention for children with craniopharyngiomas.METHODS: A panel of 6 experts, consisting of pediatric neurosurgeons and pediatric neuroradiologists, graded 30 preoperative and postoperative MRI scans according to the Puget system. Interrater reliability was calculated using Fleiss’ κ and Krippendorff’s α statistics.RESULTS: Interrater reliability in the preoperative context demonstrated moderate agreement (κ = 0.50, α = 0.51). Interrater reliability in the postoperative context was 0.27 for both methods of statistical evaluation.CONCLUSIONS: Interrater reliability for the system as defined is moderate. Slight refinements of the Puget MRI grading system, such as collapsing the 3 grades into 2, may improve its reliability, making the system more generalizable.

Modern Deep Learning (DL) networks routinely achieve classification accuracy superior to human experts, leveraging scenarios with vast amounts of training data. Community focus has now seen a shift towards the design of accurate classifiers for scenarios with limited training data. Such an example is the uncommon pediatric brain tumor, Adamantinomatous Craniopharyngioma (ACP). Recent work has demonstrated the efficacy of Transfer Learning (TL) and novel loss functions for the training of DL networks in limited data scenarios. This work describes a DL approach utilizing TL and a state-of-the-art custom loss function for predicting ACP diagnosis from radiographic data, achieving performance (CT AUPR=0.99±0.01, MRI AUPR=0.99±0.02) superior to reported human performance (0.87).

Adamantinomatous craniopharyngioma (ACP) is a rare epithelial tumor that arises in the suprasellar region. The clinical complexity associated with ACP requires unique surgical goals, making preoperative diagnosis highly desirable. Presently, definitive diagnosis of ACP requires operative tissue sampling. ACP harbors unique features on axial imaging, including cystic and calcified components. Convolutional neural networks (CNN) are computational systems that enable machines to learn operations via the assessment of examples that are task-independent, and are applied to various research fields such as computer vision and machine learning. CNNs are particularly useful with image recognition, as this preserves spatial relationships of data. CNNs have been shown to successfully identify images without a priori information of features. Inductive Transfer is a machine learning concept that is concerned with storing knowledge gained from solving a problem and applying that knowledge to a different problem. Using Inductive Transfer to overcome relatively low class populations, we leveraged the knowledge learned from several CNN models, including the Google® Inception-v4 architecture. CNNs were used to classify ACP out of a preoperative medical image dataset that also included other pathologies present in ACP differential diagnoses, such as pituitary adenoma, optic tract glioma, etc. Even baseline performance rivalled that of human performance by clinical experts. Our results suggest that CNNs can be utilized improve the reliability of pre-operative diagnosis of ACP, thereby improving clinical care.

Advances in Artificial Intelligence (AI) have rapidly become well known through the mainstream media, with myriad applications. Most commonly, AI has been leveraged for problems that possess vast amounts of sample data, as this enables AI models to better generalize patterns within the data. We hypothesized that novel computational methods could allow AI technologies to be applied to topics with more limited amounts of data, such as pediatric brain tumors. Craniopharyngioma, which has a frequency of approximately 2 in 1,000,000 children per year, requires very different treatment algorithms than other mass lesions that commonly present in the same anatomical region. As such, a prioriconfidence regarding the diagnosis of ACP would offer a distinct advantage to healthcare providers, patients, and caregivers. This makes ACP an idea model to test the application of novel AI methods. Through our collected multi-institutional dataset of CT and MRI comprising approximately 100 ACP patients controlled against 100 patients diagnosed with other suprasellar lesions, we demonstrate that state-of-the-art networks can correctly identify ACP with accuracy levels exceeding 88%, which we demonstrate to be comparable or better than current diagnostic techniques. Using concepts such as data scaling, synthetic data expansion, image augmentation, transfer-learning, and fine-tuning, this preliminary work demonstrates the reliable application of AI models towards topics with restricted volumes of data.