Usage of Deep Learning-Based Portal Dose Images for Treatment Error Detection with Transit Dosimetry

E. Uwitonzea, b, R. Lanzillottaa, c, C. Mozzid, e, L. Marinia, c, M. Avanzof, F. Lizzia, L. Marrazzod, e, g, I. Meattinid, g, S. Pallottad, e, g, G. Pirronef, A. Reticoa, C. Talamontid, e, g, A.C. Kraana aNational Institute for Nuclear Physics (INFN), Section of Pisa, Largo Bruno Pontecorvo 3, 56127 Pisa, Italy bRwanda Cancer Centre, KK 739 Street, Kicukiro, Kigali, Rwanda cUniversity of Pisa, Largo Bruno Pontecorvo 3, 56127 Pisa, Italy dUniversity of Florence, Viale Morgagni 50, 50134 Florence, Italy eNational Institute for Nuclear Physics (INFN), Section of Florence, Via Sansone 1, 50019 Sesto Fiorentino, Italy fCentro di Riferimento Oncologico di Aviano (CRO) IRCCS, Via Franco Gallini 2, 33081 Aviano, Italy gCareggi University Hospital, Largo Giovanni Alessandro Brambilla 3, 50134 Florence, Italy

Full Text PDF

In-vivo dose monitoring with electronic portal imaging devices (EPIDs) in radiotherapy can be performed by comparing a recorded EPID image with a reference expected image, either directly or by first converting it into water-equivalent dose (portal dose). We developed a deep learning model that transforms EPID images into water-equivalent dose images. An analysis framework was created, which compares the portal dose images with expectations. In this work, we test whether the framework, based on the gamma analysis, can detect changes in monitor units. We irradiated an inhomogeneous phantom using a geometric field and acquired EPID images with a planned number of monitor units, as well as with a controlled excess of monitor units. These EPID images were transformed into portal dose images using our deep learning model. Then we evaluated how the gamma passing rate varied with delivered monitor units, investigating several tolerance criteria and thresholds. The errors in monitor units could be well identified by our deep learning-based portal dose analysis framework. The choice of the gamma index threshold influenced the action level. Using a dose threshold of 10% and a tolerance of 3%/3 mm, the gamma passing rate dropped below 95% when the error in monitor units was larger than 102.4 and 102.3 for global and local gamma index calculations, respectively. For an 80% threshold, the values decreased to 102.2 for global and local calculations. When using a newly developed analysis framework, based on comparison of deep learning-based portal dose images with the gamma index, we confirmed that we can detect errors in monitor units. As a starting point for an alert system, the global gamma index analysis can be used with an 80% threshold and 3%/3 mm or stricter criterion.

DOI:10.12693/APhysPolA.148.S74 topics: portal dose (PD) images, transit dosimetry, radiotherapy, deep learning