Hayley M. Reynolds

Cell density in prostate histopathology images as a measure of tumor distribution

We have developed an automatic technique to measure cell density in high resolution histopathology images of the prostate, allowing for quantification of differences between tumour and benign regions of tissue. Haemotoxylin and Eosin (H&E) stained histopathology slides from five patients were scanned at 20x magnification and annotated by an expert pathologist. Colour deconvolution and a radial symmetry transform were used to detect cell nuclei in the images, which were processed as a set of small tiles and combined to produce global maps of cell density. Kolmogorov-Smirnov tests showed a significant difference in cell density distribution between tumour and benign regions of tissue for all images analyzed (p < 0.05), suggesting that cell density may be a useful feature for segmenting tumour in un-annotated histopathology images. ROC curves quantified the potential utility of cell density measurements in terms of specificity and sensitivity and threshold values were investigated for their classification accuracy. Motivation for this work derives from a larger study in which we aim to correlate ground truth histopathology with in-vivo multiparametric MRI (mpMRI) to validate tumour location and tumour characteristics. Specifically, cell density maps will be registered with T2-weighted MRI and ADC maps from diffusion-weighted MRI. The validated mpMRI data will then be used to parameterise a radiobiological model for designing focal radiotherapy treatment plans for prostate cancer patients.

Ensemble Prostate Tumor Classification in H&E Whole Slide Imaging via Stain Normalization and Cell Density Estimation

Classification of prostate tumor regions in digital histology images requires comparable features across datasets. Here we introduce adaptive cell density estimation and apply H&E stain normalization into a supervised classification framework to improve inter-cohort classifier robustness. The framework uses Random Forest feature selection, class-balanced training example subsampling and support vector machine SVM classification to predict the presence of high- and low-grade prostate cancer HG-PCa and LG-PCa on image tiles. Using annotated whole-slide prostate digital pathology images to train and test on two separate patient cohorts, classification performance, as measured with area under the ROC curve AUC, was 0.703 for HG-PCa and 0.705 for LG-PCa. These results improve upon previous work and demonstrate the effectiveness of cell-density and stain normalization on classification of prostate digital slides across cohorts.

Predicting prostate tumour location from multiparametric MRI using Gaussian kernel support vector machines: a preliminary study

The performance of a support vector machine (SVM) algorithm was investigated to predict prostate tumour location using multi-parametric MRI (mpMRI) data. The purpose was to obtain information of prostate tumour location for the implementation of bio-focused radiotherapy. In vivo mpMRI data were collected from 16 patients prior to radical prostatectomy. Sequences included T2-weighted imaging, diffusion-weighted imaging and dynamic contrast enhanced imaging. In vivo mpMRI was registered with ‘ground truth’ histology, using ex vivo MRI as an intermediate registration step to improve accuracy. Prostate contours were delineated by a radiation oncologist and tumours were annotated on histology by a pathologist. Five patients with minimal imaging artefacts were selected for this study. A Gaussian kernel SVM was trained and tested on different patient data subsets. Parameters were optimised using leave-oneout cross validation. Signal intensities of mpMRI were used as features and histology annotations as true labels. Prediction accuracy, as well as area under the curve (AUC) of the receiver operating characteristics (ROC) curve, were used to assess performance. Results demonstrated the prediction accuracy ranged from 70.4 to 87.1% and AUC of ROC ranged from 0.81 to 0.94. Additional investigations showed the apparent diffusion coefficient map from diffusion weighted imaging was the most important imaging modality for predicting tumour location. Future work will incorporate additional patient data into the framework to increase the sensitivity and specificity of the model, and will be extended to incorporate predictions of biological characteristics of the tumour which will be used in bio-focused radiotherapy optimisation.

Optimised robust treatment plans for prostate cancer focal brachytherapy

Abstract Focal brachytherapy is a clinical procedure that can be used to treat low-risk prostate cancer with reduced side-effects compared to conventional brachytherapy. Current practice is to man- ually plan the placement of radioactive seeds inside the prostate to achieve a desired treatment dose. Problems with the current practice are that the manual planning is time-consuming and high doses to the urethra and rectum cause undesirable side-effects. To address this problem, we have designed an optimisation algorithm that constructs treatment plans which achieve the desired dose while minimizing dose to organs at risk. We also show that these seed plans are robust to post-operative movement of the seeds within the prostate.

Performance assessment of automated tissue characterization for prostate H and E stained histopathology

Reliable automated prostate tumor detection and characterization in whole-mount histology images is sought in many applications, including post-resection tumor staging and as ground-truth data for multi-parametric MRI interpretation. In this study, an ensemble-based supervised classification algorithm for high-resolution histology images was trained on tile-based image features including histogram and gray-level co-occurrence statistics. The algorithm was assessed using different combinations of H and E prostate slides from two separate medical centers and at two different magnifications (400x and 200x), with the aim of applying tumor classification models to new data. Slides from both datasets were annotated by expert pathologists in order to identify homogeneous cancerous and non-cancerous tissue regions of interest, which were then categorized as (1) low-grade tumor (LG-PCa), including Gleason 3 and high-grade prostatic intraepithelial neoplasia (HG-PIN), (2) high-grade tumor (HG-PCa), including various Gleason 4 and 5 patterns, or (3) non-cancerous, including benign stroma and benign prostatic hyperplasia (BPH). Classification models for both LG-PCa and HG-PCa were separately trained using a support vector machine (SVM) approach, and per-tile tumor prediction maps were generated from the resulting ensembles. Results showed high sensitivity for predicting HG-PCa with an AUC up to 0.822 using training data from both medical centres, while LG-PCa showed a lower sensitivity of 0.763 with the same training data. Visual inspection of cancer probability heatmaps from 9 patients showed that 17/19 tumors were detected, and HG-PCa generally reported less false positives than LG-PCa.

Diffusion weighted and dynamic contrast enhanced MRI as an imaging biomarker for stereotactic ablative body radiotherapy (SABR) of primary renal cell carcinoma

Purpose To explore the utility of diffusion and perfusion changes in primary renal cell carcinoma (RCC) after stereotactic ablative body radiotherapy (SABR) as an early biomarker of treatment response, using diffusion weighted (DWI) and dynamic contrast enhanced (DCE) MRI. Methods Patients enrolled in a prospective pilot clinical trial received SABR for primary RCC, and had DWI and DCE MRI scheduled at baseline, 14 days and 70 days after SABR. Tumours <5cm diameter received a single fraction of 26 Gy and larger tumours received three fractions of 14 Gy. Apparent diffusion coefficient (ADC) maps were computed from DWI data and parametric and pharmacokinetic maps were fitted to the DCE data. Tumour volumes were contoured and statistics extracted. Spearman’s rank correlation coefficients were computed between MRI parameter changes versus the percentage tumour volume change from CT at 6, 12 and 24 months and the last follow-up relative to baseline CT. Results Twelve patients were eligible for DWI analysis, and a subset of ten patients for DCE MRI analysis. DCE MRI from the second follow-up MRI scan showed correlations between the change in percentage voxels with washout contrast enhancement behaviour and the change in tumour volume (ρ = 0.84, p = 0.004 at 12 month CT, ρ = 0.81, p = 0.02 at 24 month CT, and ρ = 0.89, p = 0.001 at last follow-up CT). The change in mean initial rate of enhancement and mean Ktrans at the second follow-up MRI scan were positively correlated with percent tumour volume change at the 12 month CT onwards (ρ = 0.65, p = 0.05 and ρ = 0.66, p = 0.04 at 12 month CT respectively). Changes in ADC kurtosis from histogram analysis at the first follow-up MRI scan also showed positive correlations with the percentage tumour volume change (ρ = 0.66, p = 0.02 at 12 month CT, ρ = 0.69, p = 0.02 at last follow-up CT), but these results are possibly confounded by inflammation. Conclusion DWI and DCE MRI parameters show potential as early response biomarkers after SABR for primary RCC. Further prospective validation using larger patient cohorts is warranted.

Radiobiological parameters in a tumour control probability model for prostate cancer LDR brachytherapy

To provide recommendations for the selection of radiobiological parameters for prostate cancer treatment planning. Recommendations were based on validation of the previously published values, parameter estimation and a consideration of their sensitivity within a tumour control probability (TCP) model using clinical outcomes data from low-dose-rate (LDR) brachytherapy. The proposed TCP model incorporated radiosensitivity (α) heterogeneity and a non-uniform distribution of clonogens. The clinical outcomes data included 849 prostate cancer patients treated with LDR brachytherapy at four Australian centres between 1995 and 2012. Phoenix definition of biochemical failure was used. Validation of the published values from four selected literature and parameter estimation was performed with a maximum likelihood estimation method. Each parameter was varied to evaluate the change in calculated TCP to quantify the sensitivity of the model to its radiobiological parameters. Using a previously published parameter set and a total clonogen number of 196 000 provided TCP estimates that best described the patient cohort. Fitting of all parameters with a maximum likelihood estimation was not possible. Variations in prostate TCP ranged from 0.004% to 0.67% per 1% change in each parameter. The largest variation was caused by the log-normal distribution parameters for α (mean, [Formula: see text], and standard deviation, σ α ). Based on the results using the clinical cohort data, we recommend a previously published dataset is used for future application of the TCP model with inclusion of a patient-specific, non-uniform clonogen density distribution which could be derived from multiparametric imaging. The reduction in uncertainties in these parameters will improve the confidence in using biological models for clinical radiotherapy planning.

Voxel-wise prostate cell density prediction using multiparametric magnetic resonance imaging and machine learning

Abstract Background: There are currently no methods to estimate cell density in the prostate. This study aimed to develop predictive models to estimate prostate cell density from multiparametric magnetic resonance imaging (mpMRI) data at a voxel level using machine learning techniques. Material and methods: In vivo mpMRI data were collected from 30 patients before radical prostatectomy. Sequences included T2-weighted imaging, diffusion-weighted imaging and dynamic contrast-enhanced imaging. Ground truth cell density maps were computed from histology and co-registered with mpMRI. Feature extraction and selection were performed on mpMRI data. Final models were fitted using three regression algorithms including multivariate adaptive regression spline (MARS), polynomial regression (PR) and generalised additive model (GAM). Model parameters were optimised using leave-one-out cross-validation on the training data and model performance was evaluated on test data using root mean square error (RMSE) measurements. Results: Predictive models to estimate voxel-wise prostate cell density were successfully trained and tested using the three algorithms. The best model (GAM) achieved a RMSE of 1.06 (± 0.06) × 103 cells/mm2 and a relative deviation of 13.3 ± 0.8%. Conclusion: Prostate cell density can be quantitatively estimated non-invasively from mpMRI data using high-quality co-registered data at a voxel level. These cell density predictions could be used for tissue classification, treatment response evaluation and personalised radiotherapy. Graphical Abstract

Prostate cancer focal brachytherapy: Improving treatment plan robustness using a convolved dose rate model

Abstract Low-risk prostate cancer can be treated by focal brachytherapy, in which small radioactive seeds are implanted directly into the prostate at targeted locactions. Treatment planning is complicated by post-operative displacement of the seeds from their intended location. This reduces the actual dose received by the prostate and increases the dose to surrounding tissue such as the urethra and rectum, potentially causing harmful side-effects. Current treatment planning methods do not explicitly incorporate the effect of post-operative seed displacement. To address this, the radiation dose rate function used during planning is modified to reflect displacement using convolution. This new dose rate model enables plans to be produced automatically and efficiently. Simulation experiments show that treatment plans made using the convolved dose rate function are more robust to seed displacement than those using the original unconvolved dose, preserving treatment efficacy but giving increased protection to surrounding tissue.