T. Mattfeldt et al., Prediction of prostatic cancer progression after radical prostatectomy using artificial neural networks: a feasibility study, BJU INT, 84(3), 1999, pp. 316-323
Objective To report: a methodological feasibility study in a small series o
f patients with node-negative organ-confined prostatic cancer, using artifi
cial neural networks to predict tumour progression after radical prostatect
omy and thus help to identify high-risk patients who would benefit from adj
uvant treatment.
Patients and methods A group of 20 patients with pT2N0 prostatic cancer and
postoperative tumour progression was compared with a control group of 20 p
atients with no progression, matched for age, duration of follow-up and pre
operative serum prostate-specific antigen level. Histopathological data wer
e obtained from the radical prostatectomy specimens, i.e. the Gleason score
, World Health Organisation (WHO) grade and maximum diameter of the tumour
transacts. The volume and surface area of the epithelial tumour component a
nd of the lumina of the neoplastic glands per unit tissue volume were estim
ated by morphometric methods. To predict recurrence, multilayer feedforward
networks with backpropagation (MLFF-BP), two implementations of learning v
ector quantization (LVQ), and linear discriminant analysis (LDA) were appli
ed. The ability of these models to correctly classify new cases was tested
using the 'leave-one-out' technique,
Results Progression was predicted correctly in 85% of newly presented cases
from the three routine histopathological variables alone. On the basis of
the four morphometric variables alone progression was predicted correctly i
n 93% of cases. The use of an seven variables as input data only slightly i
mproved the quality of prediction. The best results were obtained by the LV
Q networks and LDA, followed by MLFF-BP networks.
Conclusions In this methodological feasibility study, the progression of pT
2N0 prostatic cancer after radical prostatectomy could be predicted with go
od accuracy, sensitivity and specificity from routine variables or morphome
tric texture variables using artificial neural networks, These results sugg
est that: this approach should be assessed in a prospective study with more
cases.