Spectrum of errors in nodule detection and characterization using machine learning: A pictorial essay

Abstract

In academic and research settings, computer-aided nodule detection software has been shown to increase accuracy, efficiency, and throughput. However, radiologists need to be familiar with the spectrum of errors that can occur when these algorithms are employed in routine clinical settings. We review the spectrum of errors that may result from computer-aided nodule detection. In our clinical practice, we have seen errors in nodule detection, nodule localization, and nodule characterization. Each of these categories are demonstrated with illustrative cases. Through these illustrative cases, readers can be more familiar with nuances and pitfalls generated by computer-aided detection software. Although computer-aided nodule detection software is rapidly advancing, radiologists still need to thoroughly review images with mindfulness of some of the errors that can be generated by AI platforms for nodule detection.