Using Machine Learning and Radiation Dose Structured Reports to Identify the Embolization Events During Fluoroscopic Procedures

Video recordings of invasive procedures are being used for teaching and quality assurance {1,2}.  While multiple technologies facilitate video recording, manual analysis of video remains time consuming.  This has created interest in automated analysis such as recognition of key procedural segments {3}.  When pairing video data with data from the Radiation Dose Structured Reports (RDSRs), we recognized that some procedural segments such as particle embolization (embop) might be identified using the RDSR data alone.   To test this hypothesis, we trained machine learning algorithms and tested their ability to detect embop events.

Materials and Methods:

Video recordings from 19 procedures with or without embop events were obtained from our archive {4}.  Videos were reviewed by an experienced interventional radiologist and the more than 3300 RDSR events were manually labeled as +/- embop.  A subset of the labeled RDSR data was then used to train and test different machine learning models.  For training, the majority class (- embop) were purposefully undersampled.  Long Short Term Memory (LSTM), Gated Recurrent Network and Transformer models were trained with early stoppage to mitigate overfitting. The optimal probability threshold for classification of embolization events was also determined.

Results:

Initial review of 5 uterine artery embolization procedures suggested that embop segments tended to share several key attributes.  Embo_p events often included: decrease in the fluoroscopy frame rate from 7.5 to 4 frames/second;  smaller field of view;  fewer changes in table position or beam angle.  After training with data from 19 procedures, the LSTM model performed best, reaching an accuracy of 89% (Table). The LSTM model uses input, forget and output gates to control information flow and is designed to handle long-term dependencies in sequential data.

Conclusion:

Results from this small, retrospective study using only RDSR data are promising.  We expect performance will improve further as more training data becomes available.  We also expect data from other sources such as the electronic medical record (e.g. supplies, billing codes and procedure reports) as well as features extracted from the images themselves will lead to improved detection of embolization events.  We also expect other segments of fluoroscopic procedures may be detected with machine learning.