Automated detection of enteric tubes misplaced in the respiratory tract on chest radiographs using deep learning with two centre validation

Artificial intelligence in the detection of non-biological materials

“…Research has also explored the use of AI in detecting enteric tubes, though the performance of these models leaves room for improvement (Table 1 ). Mallon et al’s algorithm detected critically misplaced enteric tubes with sensitivities and specificities of 80% and 92%, respectively [ 55 ]. Other authors reported sensitivities of 100% and specificities of 76%, respectively, in identifying enteric tube malposition [ 56 ].…”

“…However, it is essential to address the elevated rate of false positives and negatives reported by these algorithms. Analysis of one model noted false positives due to ECG leads and endobronchial barium and false negatives when multiple tubes were present [ 55 ]. Figure 5 showcases class activation maps utilized to conduct failure analysis for the false positives and negatives reported.…”

“…Additionally, some models highlighted many irrelevant features, a frequent flaw of algorithms that analyze the whole image. Further, applying segmentation techniques to circumvent this issue adds complexity and room for other sources of error [ 55 , 56 ]. At present, the high number of false positives and negatives associated with these models minimizes their efficacy but highlights their potential role as a complementary tool to human readers.…”