SP55. Leveraging AI’s Potential: Real-Time Prediction Of Longitudinal Patient Development Using Imaging Data

Abstract

Purpose: Evaluating pediatric cranial growth is integral in the early identification of cranial pathology. However, clinical metrics lack sensitivity and existing statistical models can’t provide personalized predictions due to limited longitudinal pediatric datasets. We present a new method that uses deep learning to predict personalized head development, factoring in age and sex, from a single CT image or 3D photogram. Methods: We used a cross-sectional CT image dataset (N=2,020, age 0-10 years, 1,081 male, 939 female) to train a new deep-learning architecture to create age- and sex-independent patient phenotype representations from image observations. These representations are then combined with sex and future age to predict patient anatomies under the age of 10 years. We evaluated our method using an independent longitudinal CT image dataset of 61 subjects (age 2.24 ± 2.22 years, 36 male, 25 female, time between studies 1.29 ± 1.71 years). We also studied the impact of sex. Results: We obtained an error of 2.21 ± 1.42 mm predicting development at each location on the head and a volume growth error of 0.11 ± 0.09 L, significantly outperforming other state-of-the-art methods. Finally, a modification of sex information caused increased predictive errors (0.13 ± 0.10 L, p=0.01), emphasizing the role of sex in development. Conclusion: This approach is the first to create age- and sex-agnostic patient phenotype representations to predict normative development using single image information, which can be used as personalized references to identify developmental anomalies associated with pathology.

Ähnliche Arbeiten