A Stacked Generalization Chest-X-Ray-Based Framework for Mispositioned Medical Tubes and Catheters Detection

Tubes and catheters are medical devices introduced into the human body to help ill patients in critical health conditions. However, several positioning errors occur during or after the placement of such devices (Endotracheal tubes mispositioned in 10 to 20% of intubations). In addition, the delay of X-ray diagnosis after surgery can cause serious complications. Such delays are caused by the hospitals' resourcelessness or due to workload in intensive care units. The X-rays images availability (Most used diagnosis modality in intensive care units, 40% to 50%) and the presence of tubes in those images (lines are present on 33% of X-ray images) present a fertile ground to feed DCNNs training on tube error detection tasks and reduce complications. However, training and tuning one DCNN learner to resolve tube detection is time-consuming. Therefore, we propose a custom stacked generalization framework to combine wake learners with a proposed meta learner neural network architecture to resolve tube error detection tasks. The proposed framework AUC (93.84%) outperforms other related work methods with the input size of (380pixel*380pixel). Furthermore, we demonstrated the sensibility of stacked generalization to the number of base learners. Moreover, we validated the utility of input cross-validation used to form level1-metadata for the stacked generalization. Our framework can be adapted to be integrated with a CAD (computer aid decision system) for tubes error detection. The CAD can detect errors immediately after patient screening and notify radiologists to prioritize diagnosis of cases with positioning errors to adjust tubes and reduce risks significantly.