Machine learning pardigm for dynamic contrast-enhanced MRI evaluation of expanding bladder

Abstract

Delineation of the bladder under a dynamic contrast enhanced (DCE)-MRI protocol requires robust segmentation. However, this method is subject to errors due to variations in the content of fluid within the bladder, as well as presence of air and similarity of signal intensity in adjacent organs. Introduction of the contrast media into the bladder also causes signal errors due to alterations in the shape of the bladder. To circumvent such errors, and to improve the accuracy, we adapted a machine learning paradigm that utilizes the global bladder shape. The ML system first uses the combination of low level image processing tools such as filtering, and mathematical morphology as preprocessing step. We use neural network for training the network using extracted features and application of trained model on test slices to compute the delineated bladder shapes. This ML-based integrated system has an accuracy of 90.73% and time reduction of 65.2% in over manual delineation and can be used in clinical settings for IC/BPS patient care. Finally, we apply Jaccard Similarity Measure which we report to have a mean score of 0.933 (95% Confidence Interval 0.923, 0.944)

Keywords

• machine learning
• image segmentation
• hollow organs
• rim contrast enhancement