Endoscopic Image Classification Based on DWT-CM and Improved BNN for Surgical Tool Appearances

Computerized training system for endoscopic surgeries is highly desirable for improving the surgical skills of new generation surgeons. With the development of multimedia database technologies, a training system may allow a trainee to navigate a virtual operative site rendered from archived surgical images. To realize such a system, image indexing technique for efficient retrieval and video-based rendering needs to be properly developed. We present in this work a method to classify endoscopic images based on the appearances of the surgical tools. This classification can serve as an essential indexing criteria in future endoscopic image database. A neural network method is proposed for this classification, which employs an improved three-layer back-propagation neural network (BNN) combined with a texture feature extraction based on discrete wavelet transform and co-occurrence matrix. Promising results have been obtained from our simulation experiments.