MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Detecting Errors in Radiotherapy Treatment Plans Using Deep Learning

L Ma*, D Nguyen , Y Yan , A Pompos , J Tan , W Lu , R Hannan , S Jiang , UT Southwestern Medical Center, Dallas, TX

Presentations

(Sunday, 7/29/2018) 3:00 PM - 6:00 PM

Room: Exhibit Hall

Purpose: We propose to develop a deep learning based software tool to automatically detect human errors in radiotherapy (RT) treatment plans.

Methods: In order to learn the internal logical structure of RT plans, we collect all modalities of information for training the neural network, including patients’ medical history, clinical diagnosis, medical images for planning, RT prescription and planning parameters. Various machine learning methods are used to detect anomaly in a treatment plan. Certain input parameters in text format and need to be processed by natural language processing. Convolutional neural network (CNN) processes medical images like CT, PET, MRI, as well as tumor and organ contours. Supervised learning is used to infer a reasonable prescription including treatment techniques, total dose, and dose fractionation scheme from patient’s diagnosis. Unsupervised learning is used to learn the pattern of correct data so that it’s able to flag an error when the wrong combination of parameters is given.

Results: As the first step of the study, we test the power of auto-encoder, a type of unsupervised learning, to find the wrong combination of total dose, dose per fraction, number of fraction, treatment technique and modality. As a realistic error, a ‘switched’ dose per fraction and number of fraction will results in a high reconstruction error over threshold for normal data thus the error is detected.

Conclusion: This work illustrates the potential of applying deep learning in to automatic medical error detection. In the future work more deep learning methods and channels of information will be tested to completely replace manual chart checking.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email