A machine learning approach to reducing radiation exposure to the hands of the interventionalist

Monday, March 6, 2023

12:00 PM – 1:00 PM

Location: Phoenix Convention Center, North Ballroom 120 D

Poster Presenter(s)

Mohamed Aboseria, MD

Radiology Resident, PGY-2
Beth Israel Deaconess Medical Center

Disclosure information not submitted.

Author/Co-author(s)

MH

Mohamad Hotait, n/a

IS Application Developer
Beth Israel Deaconess Medical Center
NG

Nathaniel Greenbaum, MD

Anesthesiology Physician
Beth Israel Deaconess Medical Center
SH

Steven Horng, MD

Emergency Medicine Physician
Beth Israel Deaconess Medical Center
Seth Berkowitz, MD

Interventional Radiologist
Beth Israel Deaconess Medical Center

Purpose: Exposure to ionizing radiation during image guided interventional radiology procedures can place operators at risk of early side effects such as alopecia and skin erythema, and late onset side effects such as cancer {1}. One study estimated that lifetime attributable risk from occupational radiation exposure until the age of retirement for all cancers combined for male radiologists was 338 per 100,000 individuals {2}. Radiation exposure can be minimized by reducing beam-on time, strict collimation, and other ALARA principles. The hands, in particular, receive very high radiation dose as it is the closest body part of the operator to the patient {3}. Although interventionalists are trained to keep their hand out of the beam, this may be unavoidable at times, and at other times the interventionalist and technologist may be too focused on other aspects of the intervention to practice good radiation safety. We propose that a machine learning algorithm could be trained to detect the presence of the interventionalist’s hand is in the field of view and eventually serve as an alarm to the operator to adjust the field of view or their positioning which would help reduce operator hand dose.

Materials and Methods: This binary classifier is based on a DenseNet-121 architecture {4}. Fluoroscopic Images were resized to 512x512 pixels. A custom web-application was designed to integrate with PACS to facilitate image labeling. Depending on the presence of the interventionalist’s hand in the image, images were classified as hand vs no-hand. A total of 291 hand images and 424 no-hand images were used. Training took place on a high performance computing cluster within the institutional firewall {5}. The model was run with 200 epochs and a batch size of 16 images. An initial learning rate of of 0.0001 with a decay factor of 0.9 was used.

Results: The AUC obtained during training was 0.94 with an accuracy of 85%. The model was then tested on a set of 71 images and exams withheld from training. Precision, recall, and F1 score, and accuracy were 0.89, 0.76, 0.86, and 82% respectively.

Conclusion: We created a machine learning classifier capable of detecting the presence of the interventionalist’s hand in the FOV of a fluoroscopic image. Ultimately, this technology could be used to prospectively monitor the fluoroscopic video feed and alert the operator to change their hand position or collimate the field of view. The accuracy of the classifier needs to be improved for this model to be used in the field. Future directions will include gathering more training data, labeling hands with bounding boxes, and further adjustment of hyperparameters.