Computer Vision News - July 2021

3 Summary 5 The main approach of the visual evaluation is based on the intensity. As such and because of the difficulty to define subtle lesions depending on light, the STIR MRI interpretation is causing variability. Themain aimof this article is to explore an automated deep learning-based approach that would detect and quantify inflammation and is fast and reproducible. Approach and data The dataset for this experimental work comes from University College London (UCL) hospital for a total of 30 subjects. Of them, 16 were females and 14 were males. The aim of this dataset was to evaluate prediction and responsiveness using quantitative imaging biomarkers. The protocol was MRI STIR and T1-weighted turbo spin echo sequences on a 3T Philips Ingenia scanner. Fo r previous studies of quantification, manual segmentation was performed. Two readers performed such segmentationof inflammatory lesions touse in the supervised deep learning approach. The main aim of this architecture is to provide a second opinion, such as an AI advisor for the improvement of the workflow and the reduction of the workload by filtering some of the cases. In Figure 1 you can see the overall approach on the data flow, including the manual segmentation from the readers and you can read it in even more details in the original paper. Figure 1. The data workflow and the scan availability for the creation of the AI model. Towards Deep Learning- assisted Quantification ...