Prostate cancer is the second most frequently diagnosed cancer worldwide. Suspicion of prostate cancer is usually based on results from a prostate-specific antigen blood test and a digital rectal examination. However, a definite prostate cancer diagnosis can only be stated from cancer positive results from needle biopsy. This diagnostic procedure carries a risk of serious complications for the patient and has been proved to have poor sensitivity, i.e. the needle may easily miss a tumour. At present, visual guidance using transrectal ultrasound most often aids in targeting the needle positions, however, still many prostate cancer cases remain undiagnosed after the first biopsy.

Magnetic resonance imaging (MRI) has shown promising results in the differentiation of healthy and cancerous prostate tissue. To integrate as much information as possible, more studies support the use of multiparametric MRI, i.e. a set comprising both anatomical and physiological MR images of the same prostate tissue but corresponding to different acquisition conditions, thereby reflecting different tissue properties. However, the interpretation of these images is typically performed in a qualitative manner by manual visual detection and classification of abnormal features such as intensity differences and structural irregularities. A more quantitative and reproducible approach for analysis of the multiparametric MRI is desired.

This project proposes a framework for automated localisation of prostate cancer using multiparametric MRI data comprised of images from T2-weighted, diffusion-weighted, and T1-weighted dynamic contrast-enhanced MRI.

The proposed framework consists of three processing steps: In a first step, the prostate is segmented. Each voxel within the prostate is then in a second step classified either as a cancer candidate voxel or a voxel representing normal tissue, based on a set of voxel features of intensity and texture extracted from the set of multiparametric MRI data. In a third step, the set of found cancer candidate voxels is segmented into regions by means of one of two segmentation methods, LoG edge detection and watershed transform. Based on a subsequently extracted region feature, each region is classified as a cancer region or a region most probably representing normal prostate tissue. A cancer region identified at equivalent image location within more types of MR images is considered a localised tumour.

Validation of the tumour localisation results against ground truth established by statements from an expert of true tumour locations shows promising results. The best localisation performance in terms of correct localisation of true tumours with minimum of falsely localised tumours was achieved applying LoG edge detection for the region segmentation in the step of identification of cancer regions, with 9 of 11 true tumours correctly localised and a mean number of 2.67 false positive tumours. On the other hand, use of watershed transform for the region segmentation in the step of identification of cancer regions produced correct localisation for 8 of 11 true tumours and a mean number of only 1.33 falsely localised tumours.

The proposed framework for automated prostate cancer localisation using multiparametric MRI is a promising tool in the management of prostate cancer. As a diagnostic tool it could readily aid in targeting biopsy procedures. In the long term, after a few refinements and more research, automated prostate cancer localisation using multiparametric MRI may, on its own, serve as a screening tool and totally replace the need for biopsy.