End-to-end deep learning-based determination of programmable shunt valve settings from standard two-view skull radiographs

Abstract

Accurate identification of programmable shunt valve settings is crucial for managing patients with hydrocephalus, as unintended setting changes can lead to overdrainage or underdrainage of cerebrospinal fluid (CSF). Existing artificial intelligence (AI) models primarily focus on classifying shunt valve models using pre-cropped regions of interest or tangential radiographs, which limits clinical applicability. In this study, we propose a deep learning–based end-to-end framework that automatically detects and classifies Strata programmable shunt valve settings from standard two-view (anteroposterior and lateral) cranial X-rays. A YOLOv8 OBB-based detector first localizes the valve region, followed by a rule-based geometric normalization algorithm that ensures consistent orientation across projections. Subsequently, a dual-stream convolutional neural network (CNN) performs multi-view feature fusion to determine the valve’s pressure setting. Unlike prior approaches, our method eliminates the need for specialized tangential imaging or manual cropping, reducing operator dependency and supporting feasibility in routine radiographic workflows. End-to-end evaluation yielded weighted F1-scores of 0.95 on the internal test set and 0.83 on an independent external validation dataset, supporting the potential generalizability of the proposed system. These findings support the feasibility of AI-assisted verification of Strata programmable shunt valve settings using standard skull radiographs.