Multivariate modelling of prostate cancer combining magnetic resonance derived T2, diffusion, dynamic contrast-enhanced and spectroscopic parameters

Eur Radiol. 2015 May;25(5):1247-56. doi: 10.1007/s00330-014-3479-0. Epub 2015 Mar 7.

Abstract

Objectives: The objectives are determine the optimal combination of MR parameters for discriminating tumour within the prostate using linear discriminant analysis (LDA) and to compare model accuracy with that of an experienced radiologist.

Methods: Multiparameter MRIs in 24 patients before prostatectomy were acquired. Tumour outlines from whole-mount histology, T2-defined peripheral zone (PZ), and central gland (CG) were superimposed onto slice-matched parametric maps. T2, Apparent Diffusion Coefficient, initial area under the gadolinium curve, vascular parameters (K(trans),Kep,Ve), and (choline+polyamines+creatine)/citrate were compared between tumour and non-tumour tissues. Receiver operating characteristic (ROC) curves determined sensitivity and specificity at spectroscopic voxel resolution and per lesion, and LDA determined the optimal multiparametric model for identifying tumours. Accuracy was compared with an expert observer.

Results: Tumours were significantly different from PZ and CG for all parameters (all p < 0.001). Area under the ROC curve for discriminating tumour from non-tumour was significantly greater (p < 0.001) for the multiparametric model than for individual parameters; at 90 % specificity, sensitivity was 41 % (MRSI voxel resolution) and 59 % per lesion. At this specificity, an expert observer achieved 28 % and 49 % sensitivity, respectively.

Conclusion: The model was more accurate when parameters from all techniques were included and performed better than an expert observer evaluating these data.

Key points: • The combined model increases diagnostic accuracy in prostate cancer compared with individual parameters • The optimal combined model includes parameters from diffusion, spectroscopy, perfusion, and anatominal MRI • The computed model improves tumour detection compared to an expert viewing parametric maps.

Publication types

  • Research Support, Non-U.S. Gov't

MeSH terms

  • Aged
  • Contrast Media*
  • Diffusion Magnetic Resonance Imaging / methods
  • Humans
  • Image Enhancement / methods
  • Magnetic Resonance Imaging / methods*
  • Magnetic Resonance Spectroscopy / methods*
  • Male
  • Middle Aged
  • Prospective Studies
  • Prostate / pathology
  • Prostatic Neoplasms / pathology*
  • ROC Curve
  • Sensitivity and Specificity

Substances

  • Contrast Media