Yu Sun
The University of Melbourne
PhD student

Hayley Reynolds
Peter MacCallum Cancer Centre

Darren Wraith
Queensland University of Technology

Scott Williams
Radiation oncologist
Peter MacCallum Cancer Centre

Catherine Mitchell
Peter MacCallum Cancer Centre

Annette Haworth
University of Sydney

Background and Purpose

In a specific form of focal brachytherapy validated by our group [1], termed ‘bio-focused’ therapy, reliable estimation is required for both tumour location and tumour biological characteristics which include cell density, aggressiveness and hypoxia. While previous studies have found a correlation between tumour cell density shown in histology data and apparent diffusion coefficient (ADC) values computed from multiparametric MRI (mpMRI), no quantitative predictive model has been reported for predicting tumour cell density using mpMRI data at a voxel level.

The aim of this study was to investigate the performance of machine learning methods to predict tumour cell density using mpMRI data. Following our previous work on tumour location prediction, we intended to apply similar methods, along with predictions of tumour aggressiveness and the presence of hypoxia, in bio-focused brachytherapy treatment, where non-uniform doses are calculated.


In vivo mpMRI data were collected from 29 patients before radical prostatectomy. Sequences included T2-weighted (T2w), diffusion-weighted (DWI) and dynamic contrast enhanced MRI (DCE-MRI). In vivo mpMRI was registered with histology (Reynolds, 2015), from which ground truth cell density was computed using an image processing pipeline involving colour deconvolution and non-maxima suppression.

Patients were partitioned into a training group and a test group. Haralick texture features were extracted from mpMRI data. Features with high correlations were excluded. The performances of two regression methods were assessed, including linear regression (LR) and multivariate adaptive regression splines (MARS). Parameter optimisation was performed within the training group using leave-one-out cross validation. The performance of final models was assessed by relative mean squared error (rMSE) on the test group. Paired t tests were performed for the statistical significance of different methods.


Five features from mpMRI data and 17 Haralick texture features were selected. LR had an rMSE of 13.72 (±0.21) % and identified the most important mpMRI feature to a pharmacokinetic map from DCE-MRI, followed by ADC. MARS showed marginal improvement in rMSE (0.02%, p<0.01) over LR. Both methods gave consistent results for feature relative importance.


We developed a model for prostate tumour cell density prediction at a voxel level from mpMRI using regression methods and histology data for validation. Future work will investigate other non-linear regression methods and incorporate the prediction of tumour aggressiveness and hypoxia for biological optimisations in brachytherapy treatment planning.


[1] Haworth, A. et al. (2013). Validation of a radiobiological model for low-dose-rate prostate boost focal therapy treatment planning. Brachytherapy, 12(6), 628-636.

[2] Reynolds, H. et al. (2015). Development of a registration framework to validate MRI with histology for prostate focal therapy. Medical Physics, 42(12), 7078-7089.

← Back to all abstracts