Computer Vision News - May 2022

49 the PI-RADS Assistant protocol. Lesion position was detected by comparing the segmentation labels’ position to determine lesion restriction. As our current dataset includes T2 and ADC sequences only ( The ProstateX dataset ) , with limited resolution, we expect to improve performance by introducing other sequences with higher resolution . Increasing the dataset size is also expected to have an effect on performance. As described previously, this tool can be expanded to other risk-scoring methods: BI-RADS (for breast cancer), TI-RADS (for thyroid cancer) , etc. Adjustments to the different imaging modalities need to be made, but the concept remains similar - automating region detection and calculation, to save scan reviewing time. Implementing AI methodology in these fields will significantly reduce radiologist workload and improve patient care. As AI is slowly penetrating various use-cases in healthcare, cancer screening should benefit as well. Read more about AI in urology. by the radiologist to assess cancer risk and calculate the PI-RADS score faster and in a robust manner . There are several steps for the development of the PI-RADS assistant. Initially, image preprocessing was conducted to provide adequate input to the system. T2 and ADC image protocols undergo resolution equalization and normalization. Also, gentle registration of the protocols is conducted to compensate for patient movement throughout the MRI scan. Using deep learning (DL) algorithms , the processed T2 sequence image data is passedthrougha U-Netarchitectureneural network , whose output is the segmented regions: whole prostate, transition zone (TZ), and peripheral zone (PZ). A separate network with similar architecture is used for lesion segmentation, utilizing both T2 and ADC sequences. Standard methods are used for calculation of the parameters relevant for the PI- RADS score. Size and volume are retrieved directly from the segmentation mask. Edge smoothness is calculated by the gradient along the mask’s edge, and it also is fitted to an oval shape and results in an ovalness score. Mean intensity and standard deviation is calculated within the lesion in all sequences to support the scoring