Optimizing image hanging protocols is a crucial measure to improve radiologists’ workflows. In this study, we developed a fully- automated image-hanging tool for upper abdominal MRI, which can replace the manual work of image preparation before radiological diagnosis, significantly improving work efficiency and reducing workload.
The core technique to achieve automated image hanging is MR sequence and phase discrimination. Our model can recognize thirteen different MR sequences and phases, with an average accuracy of 99.1%. These thirteen categories cover common sequences and phases of abdominal MRI, including hepatocyte-specific contrast agent MRI, ensuring the model’s broad applicability. Wang et al.[7] reported a dense feature fusion neural network for automatic sequence and phase discrimination of liver MRI and proposed that it could be used to optimize hanging protocols. However, their model was based on a multiserial information fusion approach, which leads to a decrease in performance when there are missing or added sequences. In contrast, our model’s classification of each sequence is independent, ensuring its high stability for different sequence compositions.
Dercle et al.[8] developed a machine-learning algorithm to differentiate five contrast-enhancement phases of liver CT scans. However, their approach requires radiologists to draw regions of interest around the abdominal aorta and portal vein, which limits its practical application. Our goal is not only to achieve sequence identification but also to build a complete pipeline for assisting radiologists in image preparing. Therefore, we paid extra attention to the feasibility of the model in clinical practice. To address issues n actual clinical settings, we incorporated multiple algorithms or rules into the model. For instanace, a set of MRI images may contain many sequences that are not necessary to view, such as locating sequences, sequences downsampled from other sequences for printing, and sequences acquired intermediately in the generation of other sequences, which may interfere with subsequent sequence recognition. To resolve this issue, we established some rules based on image orientation and image features to filter out these images.
In a real clinical setting, there may also be situations where a patient has more than one sequence or phase of the same type, such as using multiple arterial phase acquisition approaches. It makes the model necessary to have the ability to find the optimal image serials among multiple sequences. Our solution is to calculate a matching score for each candidate sequence through a regression model, compairing its similarity to the gold standard, and the sequence with the highest score is regarded as the optimum sequence.
To evaluate the performance of the full model in a real clinical scenario, we conducted a one-month clinical evaluation in ** Hospital. The results showed that the model successfully processed 98.5% of the target MRI scans, with 84.6% achieving 100% accuracy. It indicates that with the assistance of AI, one-click image hanging is achievable in most cases. Manual correction was needed only in a few cases, and such correction usually involved only one sequence or phase, and rarely involved more than one sequence or phase. It indicates that even if the model makes errors, modifications based on the model’s results can reduce the manual work of radiologists.
To further optimize the model, we investigated possible causes of model errors and attempted to find solutions. For example, we found that the model was prone to misclassify earlier PVP when the phases were between standard LAP and PVP (26 errors in PVP are similar to this). We are currently collecting such cases to retrain the model. Some technical problems have been corrected by adjusting and optimizing the algorithm.
To investigate the model’s scope of clinical application, we used multivariate logistic regression to analyse the factors related to model errors. The contrast agents, MR devices, and liver backgrounds were not statistically significant, indicating that our model had good stability in the explored scenarios. The reasons for choosing these three factors are as follows: first, dynamic enhanced MRI with hepatobiliary contrast and extracellular contrast has significantly different enhancement patterns of solid organs and major vessels[9, 10]. Second, imaging protocols and parameters vary between MR devices, which can be a challenge. Third, advaned chronic liver disease often leads to a series of abnormalities and circulatory changes in the abdomen[11, 12], which are likely to affect phase identification.
The model exhibits good accuracy and stability, resulting in significant improvement in efficiency. In each evaluated case, AI-assisted image hanging consumed less time than manual operation. On average, the AI tool saved 96 seconds per case for radiologists. The model eliminates the two time-consuming steps of loading and rearranging images compared to hanging images on a conventional PACS. With its assistance, doctors only need to read all the images and make minor modifications if needed.
However, our study has some limitations. First, although the model was built using multicentre data, the clinical evaluation was performed at only one centre. Therefore, it remains to be fully validated on a large multicentre dataset involving more brands of PACS, MR scanners, scanning protocols, and various diseases. Second, other than chronic liver disease, some other diseases or abnormalities in the abdomen may also affect the model’s performance, such as nephrectomy, splenectomy and vascular disease. In future research, these factors will be taken into account.