In this pilot study, we tested whether it is possible to apply neural netwo
rk-based diagnostics on bladder washings to detect urothelial cell carcinom
a of the bladder. Eighty-five bladder wash (BW) samples were chosen at rand
om from our own database. Cystoscopy, histology, and follow-up data concern
ing tumor recurrence were available. Each slide was scanned by the neural n
etwork-based digitized cell image system. The neural network-based diagnosi
s (NNBD) was based on 128 digitized cell images provided by the system. The
light microscopic diagnosis (LMD) was rendered by an experienced cytopatho
logist using the same terminology, ie., negative, low-grade tumor and high-
grade tumor. Finally, an automatic QUANTICYT analysis was performed on the
same material, with as classification low, intermediate, and high risk. The
sensitivity for diagnosing a histologically confirmed tumor was for NNBD 9
2%, for LMD 50%, and for QUANTICYT 69%.
For the three methods, receiver operating characteristic (ROC) curves were
made for the thresholds low grade/intermediate risk and high grade/high ris
k. For the prediction of a positive cystoscopy, the highest area under the
curve (AUC) was found for NNBD, being 0.71. The AUC for LMD was 0.58. QUANT
ICYT analysis had the highest AUC value (0.62) for predicting tumor recurre
nce after a negative cystoscopy, with a lower value for NNBD (0.50). These
findings indicate that neural network-based diagnosis of bladder washing sa
mples is highly promising. (C) 2000 Wiley-Liss, Inc.