development of artificial intelligence as a conversion tool for cine electronic portal imaging device images to radiotherapy dosimetry: preliminary study

Introduction: this research is a preliminary study of the development of artificial intelligence (ai) as a conversion tool from the pixel value of cine a-si 1000 electronic portal imaging device (epid) images to dose. it also investigates the relationship between the monitor unit (mu), dose rate, number of frames, and beam profile of electronic portal imaging device (epid) images to facilitate further mathematical correction that must be added to create accurate dosimetry by cine epid images. material and methods: homogeneous and inhomogeneous phantom was irradiated in a linear accelerator (linac) 6 mv with different techniques, field size, and phantom thickness. the cine a-si 1000 epid images were taken and compared to dose distribution data derived from the eclipse treatment planning system (tps) at source axis distance 100 cm or isocenter field. the ai model training process begins with the augmentation of epid and tps images from homogeneous phantom so that 1152 images are obtained. these images are then split randomly into training and testing data 7:3, and validation is done using gamma index 3%/3mm. results: an ai model based on convolutional neural network (cnn) with 6 layers has been successfully created that can convert epid pixel values into dose distribution without any mathematical correction. the best results from validation with a gamma index of 3%/3mm compared to tps calculations reached 92.40% ±28.14%. conclusion: an ai model has been successfully created that can convert epid pixel values into dose distribution but need improvement by considering the characteristics contained in the epid image and the number of datasets.