Detection of correct placement and presence of tubes on chest x-rays using Efficient Net

Abstract

In recent years the sudden rise of data in healthcare has gained huge attraction towards artificial intelligence. (AI), Powerful algorithms can detect essential features of medical data and assist doctors with the fast and accurate diagnosis of different diseases. This research aims to provide an Automatic Medical Tube Detection System (AMTD) to detect mispositioned tubes on chest Xrays during intubation. Intubation is the process of inserting tubes into patient bodies. Depending upon the patient, different types of intubations are necessary. An endotracheal tube is inserted into the patient's windpipe through the nose or mouth to help the patient breathe. The nasogastric tube is used to gain access to the stomach for feeding purposes, and the central venous catheter is inserted into a primary vein for medicinal purposes. Medical professionals use clinical protocols to insert catheters and tubes, but it is a complex and error-prone procedure, especially when hospitals are overcrowded. Any misplacement in the tubes may cause severe complications. So, early detection of wrongly positioned tubes using chest x-rays will help to save patient lives. The main objective of this research is to propose automatic medical tube detection (AMTD) system using transfer learning with CNN-based Efficient Net architecture. Our proposed methodology consists of two modules, i.e., classification of presence and wrongly positioned tubes using Efficient Net B7 with auxiliary connection and quantization of trained weights to floating-point 16 (FP16). The proposed technique uses an Auxiliary connection to prevent vanishing gradient problems and to increase the model performance. Furthermore, the trained model is extended to the light-weighted medical tube detection (LMTD) system to support edge devices. Edge devices often have limited memory and computational power. By applying post-training FP16 quantization, we have reduced the size of the LMTD model by 2X compared to the FP32 AMTD system and supported by CPUs/GPUs with no accuracy loss. In order to evaluate the performance of both AMTD and LMTD models, Experimentation was done on five different randomly split datasets with a ratio of 80:20 train and validation set, respectively. As a result, we have achieved an average Area under the ROC curve (AUC) of 0.964% for the AMTD model. The LMTD also achieves an AUC of 0.962%.