Automatic Diagnosis Framework for Catheters and Tubes Semantic Segmentation and Placement Errors Detection

Automatic diagnosis systems are highly required in daily healthcare work. Working in such a sensitive environment with a high workload and stress induces a high possibility of human errors. The tubes and catheters positioning task are one of the medical operations where mistakes are risky and cause crucial complications if not detected early. The process of tube placement consists of inserting a type of tube for a patient. Then, the patient is screened to diagnose the position of the installed tube. The chest X-Ray image must wait for a radiologist to validate the diagnosis. The delay or possible diagnosis errors can cause more complications. Through this work we propose a diagnosis validation framework for in-time error detection. The framework processes the chest X-Ray just after the tube is inserted and outputs a segmentation mask associated with the classification values for possible errors. Our proposed framework is based on a customized U-net and shows a competitive segmentation result with a dice coefficient of 94,5%. The proposed framework is optimized to be deployed on production devices with 75% training parameters less compared to the original U-Net model version.