Medical Image Thresholding Using Online Trained Neural Networks

Abstract

AbstractMedical images are used mainly in the diagnosing process and as an aid in determining correct treatment. Therefore, the process of segmenting different regions of interests (ROIs) within the medical images is considered a critical one. When provided with a segment with high segmentation accuracy, the physician can easily detect the problem and determine the best treatment. In this paper, a neural network retrained on-line is proposed to automatically segment medical images using a global threshold. The network is initially trained off-line using a set of features extracted from a set of randomly selected training images, along with their best thresholds, as targets for the neural network. The features are extracted using Seeded Up Robust Feature (SURF) technique from a rectangle around the ROI. This network continues training on-line as new images arrive, based on a feedback correction done by the clinician to the segmented image. This process is repeated multiple times to verify the generalization ability of the network.KeywordsNeural networkSeeded Up Robust Feature (SURF)