Content-based retrieval in picture archiving and communication systems.

A COntent-Based Retrieval Architecture (COBRA) for picture archiving and communication systems (PACS) is introduced. COBRA improves the diagnosis, research, and training capabilities of PACS systems by adding retrieval by content features to those systems. COBRA is an open architecture based on widely used health care and technology standards. In addition to regular PACS components, COBRA includes additional components to handle representation, storage, and content-based similarity retrieval. Within COBRA, an anatomy classification algorithm is introduced to automatically classify PACS studies based on their anatomy. Such a classification allows the use of different segmentation and image-processing algorithms for different anatomies. COBRA uses primitive retrieval criteria such as color, texture, shape, and more complex criteria including object-based spatial relations and regions of interest. A prototype content-based retrieval system for MR brain images was developed to illustrate the concepts introduced in COBRA.

Edge detection in medical images using a genetic algorithm.

An algorithm is developed that detects well-localized, unfragmented, thin edges in medical images based on optimization of edge configurations using a genetic algorithm (GA). Several enhancements were added to improve the performance of the algorithm over a traditional GA. The edge map is split into connected subregions to reduce the solution space and simplify the problem. The edge-map is then optimized in parallel using incorporated genetic operators that perform transforms on edge structures. Adaptation is used to control operator probabilities based on their participation. The GA was compared to the simulated annealing (SA) approach using ideal and actual medical images from different modalities including magnetic resonance imaging (MRI), computed tomography (CT), and ultrasound. Quantitative comparisons were provided based on the Pratt figure of merit and on the cost-function minimization. The detected edges were thin, continuous, and well localized. Most of the basic edge features were detected. Results for different medical image modalities are promising and encourage further investigation to improve the accuracy and experiment with different cost functions and genetic operators.

Detection of suspected malignant patterns in three-dimensional magnetic resonance breast images.

In this article, a Boolean Neural Network (BNN) is used for the detection of suspected malignant regions in 3D breast magnetic resonance (MR) images. The BNN is characterized by fast learning and classification, guaranteed convergence, and simple, integer weight calculations. The BNN learning algorithm is incremental, which allows the addition and deletion of training patterns without unlearning those already learned. The incremental learning algorithm automatically reduces the training set and trains the network only with those examples estimated to be useful. The architecture is suitable for parallel hardware implementation using available Very Large Scale Integration (VLSI) technology. The BNN was trained by using a set of malignant, benign, and false-positive patterns, extracted by experts, from selected MR studies, by using an incremental learning algorithm. After training, the network was tested by means of a consistency checking test, cross validation techniques, and patterns from actual MR breast images. During the consistency test, the BNN was tested by using the same patterns used for training. The BNN classification accuracy in this case was 99.75%, proving the ability of the BNN to select useful patterns from the training set. Then, a leave one out cross-validation (LOOCV) test was done by using patterns from the training set and the classification accuracy was 90%. Next, an extended training set was created by shifting the original patterns in different directions. A cross-validation test was then performed by dividing the set of patterns into a training and a test set. Classification accuracy was compared to the nearest neighbor classifier. Results showed that the BNN achieved an average of 77% classification accuracy while requiring only 34% of the original training set. On the other hand, the nearest neighbor classifier achieved an accuracy of 57.9% while retaining the whole training set. Another test using actual MR slices different from the training set was done and results compared favorably to a radiologist's findings. Test results show the BNN's capability to detect suspected malignant regions in 3D MR images of the breast. The proposed BNN architecture can save the radiologist a great deal of time browsing MR slices searching for suspected malignancies.