Abstract
Prostate cancer (PCa) is among the most frequently diagnosed cancers in men and is the second leading cause of cancer-related deaths, following lung cancer. Traditionally, prostate cancer diagnosis relies on transrectal ultrasound (TRUS)-guided biopsy. However, this method has limitations, including a risk of missing malignant regions and potential complications such as rectal bleeding. Moreover, over half of patients who undergo systematic biopsy procedures eventually require a radical prostatectomy (RP),highlighting the need for more accurate diagnostic techniques.To address these challenges, researchers have worked extensively to improve diagnostic accuracy using ultrasound (US) and magnetic resonance imaging (MRI). The introduction of multiparametric imaging has enhanced diagnostic performance. However, these techniques still face limitations, particularly in tissue differentiation and localization. One way to assess these limitations is through the use of tissue-mimicking material (TMM) phantoms, which simulate biological tissues with known mechanical and acoustic properties. This allows for controlled evaluation of imaging modalities and identification of diagnostic weaknesses.
In this study, a multiparametric ultrasound (mpUS) phantom was developed using polyvinyl alcohol (PVA) cryogel, a material known for its tunable acoustic and mechanical properties, closely resembling those of the prostate gland, cancerous tissues, blood vessels, and surrounding soft tissues. Various additives were incorporated to enhance specific properties: silicon carbide (SiC) to increase ultrasound backscatter, aluminum oxide (Al₂O₃) to improve attenuation coefficient, glycerol to adjust the speed of sound, and benzalkonium chloride (BC) to inhibit bacterial growth.
The acoustic properties—speed of sound, attenuation coefficient, and acoustic impedance—were measured using time-of-flight techniques. Mechanical properties were evaluated via an indentation test to assess elasticity in kilopascals (kPa). The phantom was then scanned using B-mode ultrasound, shear wave elastography (SWE), and Doppler ultrasound. Results demonstrated that the PVA-based phantom closely simulated human prostate tissues in both appearance and behaviour. The B-mode and SWE images were visually consistent with human scans, and Doppler imaging successfully reproduced pulsatile flow patterns resembling those of prostatic micro-vessels.
To assess the impact of phantom composition on imaging performance, an alternative phantom made from agar-based material (per IEC standards) was developed. Replacing PVA with this stiffer material led to increased Doppler velocity signals and slight variations in SWE elasticity, indicating sensitivity of imaging outcomes to tissue stiffness.
In the artificial intelligence (AI) component of the study, data were collected from 50 patients, including B-mode and SWE images confirmed by postoperative histopathological analysis. Each image contained clearly identifiable true positive and true negative regions, with some also presenting false positives and false negatives. An additional 12 patient cases consisted exclusively of false positive and false negative samples to enhance model evaluation.
A total of 94 texture features were extracted from six regions of interest (ROIs): one from B-mode and five from reconstructed SWE images. These features were derived from texture analysis techniques including the Gray-Level Co-Occurrence Matrix (GLCM), Gray-Level Run Length Matrix (GLRLM), Gray-Level Size Zone Matrix (GLSZM), and Gray-Level Distance Zone Matrix (GLDZM).
Several machine learning (ML) models—Random Forest, k-Nearest Neighbours (KNN), Support Vector Machine (SVM), Logistic Regression, and Naive Bayes—were trained to classify normal vs. malignant tissue. Model training and evaluation were conducted using five-fold cross-validation. Feature selection was performed using statistical tests and the LASSO (Least Absolute Shrinkage and Selection Operator) method to identify the most predictive features.
The results indicated that B-mode image texture features alone did not significantly differentiate between benign and malignant tissues. However, several features extracted from SWE images—especially mean intensity, contrast, and long-run emphasis—consistently demonstrated strong discriminatory power. Among all ROIs, reconstructed gray-scale SWE images offered superior performance in distinguishing prostate cancer regions.
| Date of Award | 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Benjie Tang (Supervisor) & Cheng Wei (Supervisor) |
Cite this
- Standard