A Chest X-Ray Private Dataset for Classification Models Evaluation on the Task of Endotracheal Tubes Position Diagnosis

Tracheal intubation is a medical operation for patient lung ventilation support in emergency care units. The act involves introducing a plastic tube, the Endotracheal tube, into the patient airway system. The tube must be at a precise position between the carina and vocal cords to prevent the potential risk of an adjustment delay toward an incorrect placement. A chest x-ray screening can show the error if taken after the tube insertion. That requires doctors to promptly diagnose the issue, which is not possible in overloaded work environments. A machine learning model can reduce the risk by notifying the suspect patient who needs immediate tube adjustment. We propose by this work a procedure to test the effectiveness of the tube's position diagnosis in a work environment by testing the models at an internal private dataset. We tested our method at the Ibn Sina Hospital in Rabat. We collected and annotated a private dataset with a radiologist at the emergency service. Furthermore, we trained two models, one for endotracheal X-ray image extraction from the hospital archive system and a model for tube placement diagnosis. We also customized an application for the private dataset image labeling. Finally, we proved the possible generalization of the diagnosis model by the test on the private dataset.