Prostate Cancer Aggressiveness Prediction Using CT Images

Abstract

Prostate Cancer (PCa) is mostly asymptomatic at an early stage and often painless requiring active surveillance screening. Transrectal Ultrasound Guided Biopsy (TRUS) is the principal method to diagnose PCa following a histological examination by observing cell pattern irregularities and assigning the Gleason Score (GS) according to the recommended guidelines. This procedure presents sampling errors and, being invasive may cause complications to the patients. External Beam Radiotherapy Treatment (EBRT) is presented as curative option for localised and locally advanced disease, as a palliative option for metastatic low-volume disease or after prostatectomy for prostate bed and pelvic nodes salvage. In the EBRT worflow a Computed Tomography (CT) scan is performed as the basis for dose calculations and volume delineations. In this work, we evaluated the use of data-characterization algorithms (radiomics) from CT images for PCa aggressiveness assessment. The fundamental motivation relies on the wide availability of CT images and the need to provide tools to assess EBRT effectiveness. We used Pyradiomics and Local Image Features Extraction (LIFEx) to extract features and search for a radiomic signature within CT images. Finnaly, when applying Principal Component Analysis (PCA) to the features, we were able to show promising results.

Keywords: classification; computed tomography; prostate cancer; radiomic features; risk stratification.

Figure 1

Stratification by Risk Group (RG). Adapted from [5].

Figure 2

Volumes of interest for prostate adenocarcinoma treatment planning. In orange the ctv; in red the ptv; in green the rectum; the bladder in dark blue and in light blue the femoral heads. Adapted from Gregoire et al. [7].

Figure 3

3D Slicer Interface. Visualization of image series, OARs and CTV. Top-left: Axial view. Top-right: Volumetric reconstruction. Bottom-left: Coronal view. Bottom-right: Sagittal view.

Figure 4

LIFEx Interface. Visualization of image series and ctv in blue.

Figure 5

Radiomics Pipeline.

Figure 6

Hierarchical cluster heatmap for pyradiomics features.

Figure 7

Hierarchical cluster heatmap for LIFEx features.

Figure 8

AUROC Values for multiple PCA variations and number of principal components.