Probabilistic Neural Network Analysis of Quantitative Nuclear Features in Predicting the Risk of Cancer Recurrence at Different Follow-up Times P.Petalas 1* , P. Spyridonos 1 , D. Glotsos 1 , D. Cavouras 2 , P.Ravazoula 3 , G. Nikiforidis 1 1. Computer Laboratory, School of Medicine, University of Patras, Rio, Patras 265 00, Greece 2. Department of Medical Instrumentation Technology, Technological Education Institution of Athens, Ag. Spyridonos Street, Aigaleo, 122 10, Athens, Greece 3. Department of Pathology, University Hospital, Rio, Patras, 265 00, Greece *petalas.p@med.upatras.gr Abstract In this study, we explore the prognostic significance of automatically generated features from biopsies of superficial transitional cell carcinomas (TCCs) of urinary bladder in predicting the time interval in which the tumor is more likely to recur. Clinical material comprised 73 patients diagnosed with superficial TCC and followed-up for at least 60 months. Patients’ data set was separated into three prognostic groups in respect to time interval in which the tumor may recur. For each case, 40 descriptive quantitative features related to nuclear characteristics were generated. The prognostic value of the estimated features was analyzed by means of a Probabilistic Neural Network (PNN) classifier. To find best vector combination leading to the smallest classification error, an exhaustive search procedure in feature space was utilized. Classifier performance was evaluated by means of the leave-one-out method. Throughout the analysis of the prognostic feature combinations, four features, two describing nuclear texture, and two related to shape distribution of nuclei in the sample, were identified as the important markers for patients’ outcome prediction. The overall predictive accuracy was 73%. Inter-mediate and long-term recurrent cases were identified with an accuracy of 76%. For the short-term group the predictive accuracy was 63%. The detection of recurrences in certain future time seems feasible by analyzing the prognostic information of nuclear features and incorporating the PNN model. The improvement in prognostic ability may be clinically important for patient follow-up, and therapeutic treatment. 1. Introduction Urinary bladder cancer is now the fourth most common malignancy in the western male population [1]. The majority of bladder tumors are transitional cell carcinomas (TCCs). Over 60% of patients affected with superficial tumors will have one or more recurrences after initial treatment and about 15% will progress and eventually die of the disease [2]. In managing patients with bladder cancer, one of the most significant problems for the clinician is the prediction of tumor recurrence [3]. Currently, the risk of recurrence has been investigated on the basis of various clinical and pathological data [4]. Prognostic factors such as patient age, sex, stage and grade have been analyzed by means of an Artificial Neural Network (ANN) with respect to disease recurrence. The ANN classified non-recurrent tumors with 55% accuracy and recurrent tumors with 76% [5]. The use of Kaplan-Meyer analysis indicated that variables generated by cytometry can provide a cut-off value (p<0.00001) for the discrimination between low and high risks tumor recurrence [6]. In a previous work on TCCs recurrence, we have proposed a prognostic system based on a combination of histopathological and nuclear features processed by an ANN classifier. The system has performed a significant prognostic assessment. The predictive accuracy for recurrent tumors was 74.5% and for non-recurrent tumors 71.1% [7]. In this study, we explore the prognostic significance of automatically generated features from biopsies of superficial urinary bladder tumors in predicting the time interval in which the tumor is more likely to recur. The prognostic value of the estimated features was analyzed by means of a Probabilistic Neural Network (PNN) classifier. Defining the risk of recurrence at certain time intervals will improve patients’ follow-up and facilitate treatment selection. To the best of our knowledge, there is no previous research work in predicting how long after the initial tumor resection the cancer will recur. 2.Materials and Methods 2.1 Patients data set The clinical material comprised 73 patients (cases) diagnosed with superficial TCC (pTa, pT1,) and followed-up during the period 1991-1998 at University Hospital of Patras Greece. The follow-up period was at