PROJECT TITLE :

Systematic Clinical Evaluation of a Deep Learning Method for Medical Image Segmentation Radiosurgery Application

ABSTRACT:

We conduct an in-depth analysis of a Deep Learning model by using it to segment three-dimensional medical images. We address the shortcomings of manual segmentation, which include high inter-rater contouring variability and time consumption during the process of contouring, by developing our own model. The careful and detailed analysis, which could be further generalized on other medical image segmentation tasks, is the primary extension that is made in comparison to the previous evaluations. First, we look at how changes in the detection agreement between different raters have occurred. We are able to demonstrate that the model reduces the number of disagreements in detection by 48% (p 0.05). Second, we demonstrate that the model enhances the inter-rater contouring agreement, which went from 0.845 to 0.871 surface Dice Score (p 0.05) when using the model. Thirdly, we demonstrate that by using the model, the delineation process can be sped up between 1.6 and 2.0 times more quickly (p 0.05). In conclusion, we design the framework of the clinical experiment in such a way as to either exclude or estimate the evaluation biases; this helps to ensure that the results' significance is maintained. In addition to the clinical evaluation, we also discuss our intuitions and ideas for the practical application of developing an effective DL-based model for the segmentation of 3D medical images.