Radiographic Assessment of CVC Malpositioning: How can AI best support clinicians?

Lasse Hansen, Malte Sieren, Malte Hobe, Axel Saalbach, Heinrich Schulz, Jörg Barkhausen, Mattias P Heinrich

Keywords: Chest X-Ray, CVC Malpositioning, Explainable AI

Abstract:

The malpositioning of central venous catheters (CVCs) is a common technical complication that is usually diagnosed on post-procedure chest X-rays (CXRs). If the misplaced CVC remains undetected, it can lead to serious health consequences for the patient. Interpreting CXRs at a large scale in everyday clinical practice is time consuming and can create delays in the repositioning of the CVC. A computer-assisted assessment of post-procedure CXRs can help to prioritise cases and reduce human errors in stressful situations by objectifying decisions. However, final assessments must always be made by the clinicians, which is why the algorithmic support needs to be comprehensible.
Since AI models are not yet able to detect catheter maplpositons with highest accuracy, the focus must be on efficient support in everyday clinical practice. In this work, we evaluate three different AI models, particularly with regard to the relationship between classification accuracy and the degree of explainability. Our results show how helpful it is to incorporate explicit clinical knowledge into deep learning-based models and give us promising research directions for a planned large scale patient study.

Paper Type: validation/application paper

Primary Subject Area: Application: Radiology

Secondary Subject Area: Interpretability and Explainable AI

Paper Status: original work, not submitted yet

Source Code Url: Both networks (segmentation and classification), are based on the following public repositories: https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/efficientnet.py, https://github.com/qubvel/segmentation_models.pytorch

Data Set Url: https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification, the UKSH dataset can currently not be published due to legal regularisations

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.