Artificial neural network analysis (ANNA) of prostatic transrectal ultrasound

BACKGROUND. Our purpose was to determine the diagnostic potential of a new, computerized method of interpreting transrectal ultrasound (TRUS) informat ion by artificial neural network analysis (ANNA). This method was developed to resolve the current dilemma of visual differentiation between benign an d malignant tissue on TRUS. To train and objectively evaluate ANNA, a new p recise method of computerized virtual correlation of preoperative ultrasoun d findings and radical prostatectomy histopathology was devised. After trai ning with this pathologically confirmed digitized TRUS information, ANNA wa s tested in a blinded study. METHODS. Following radical prostatectomy, 289 pathology whole-mount section s of 61 patients were correlated digitally with the corresponding TRUS slic es. Specific selection of TRUS areas unequivocally identified on the correl ated digitized pathohistology resulted in 553 pathology-confirmed represent ations (samples). Of these, 53 were used for training and 500 were subjecte d to blind analysis by ANNA. RESULTS. ANNA classified 378 (99%) of the 381 benign pathology-confirmed sa mples correctly as benign. The false-positive rate was 1% (n = 3). Of the 1 19 pathology-confirmed malignant samples, 94 (79%) were classified correctl y; 25(21%) were falsely classified as normal. Out of all 119 cancers, ANNA classified 60 (71%)bf the hypoechoic cancers as malignant and 24 (29%) as b enign. Surprisingly, 34(97%) of the isoechoic cancers were correctly classi fied by ANNA, missing only one sample. CONCLUSIONS. The introduction of ANNA enhanced the accuracy of TRUS prostat e cancer identification. Although not all malignant areas were detected, ca ncer was detected in each patient. The ability to detect isoechoic cancerou s lesions appears to be the essential innovation over conventional TRUS int erpretation. (C) 1999 Wiley-Liss, Inc.