The effectiveness of automatic laparoscopic diagnostics of liver pathology using different methods of digital images classification
Сomputer automatic diagnostic (CAD)/classification of videoimages is actual for laparoscopic surgery. Such CAD is supposed to explore intraoperatively for support surgeon decisions. Aim: to evaluate the effectiveness of the CAD systems developed on the basis of two classifiers -HAAR features cascade and AdaBoost for the detection of cirrhotic and metastatic damages of the liver. Materials and methods. The development of CAD was based on training of HAAR features cascade and AdaBoost classifiers
... daBoost classifiers with images/frames, which have been cropped out from video gained in the course of laparoscopic diagnostics. RGB frames which were gamma-corrected and converted into HSV have been used for training. Also descriptors were extracted from images with the modified method of Local Binary Pattern (LBT), which includes data on color characteristics ("modified color LBT" -MCLBT) and textural ones for AdaBoost classifier training. 1000 positive images along with 500 negative ones of both types of pathology were used for training. After cessation of training the tests were performed with the aim of the estimation of effectiveness of recognition. Test session images were different from those ones which have been used for training of the classifier. Test control sessions were performed with trained classifiers with 319 frames containing cirrhotic and 253 frames with metastatic deteriorations in liver tissue. 365 frames with the absence of mentioned pathology were used as a control group -practically healthy liver state. Results. Classification of test video-images revealed that the highest recall for cirrhosis diagnostics was achieved after training of AdaBoost with MCLBT descriptors extracted from HSV images -0.655, and in case for metastatic damages diagnostics -for MCLBT gained from RGB images -0.925. Hence developed AdaBoost based CAD system achieved 69.0 % correct classification rate (accuracy) for cirrhotic and 92.7 % for metastatic images. The accuracy of Haar features classifier was highest in case of metastatic foci identification and achieved 0.701 (RGB) -0.717 (HSV) values. Conclusions. Haar features based cascade classifier turned to be less effective when compared with AdaBoost classifier trained with MCLBT descriptors. Metastatic foci are better diagnosed when compared with cirrhotic liver deterioration with the explored approaches to digital images classification.