Abstract
Prostate cancer is the second most common cancer in men world-wide, with approximately 174,650 new cases diagnosed in 2019 in the U.S. However, prognosis is relatively good given sufficiently early detection during the non-metastatic stage, motivating the need for fast and reliable cancer screening methods. Diffusion weighted imaging is a magnetic resonance imaging technique that is gaining traction as a noninvasive method for cancer screening. In 2013, a new form of diffusion weighted imaging called correlated diffusion imaging (CDI) was introduced as a potential candidate modality for building computer-aided clinical decision support systems. We perform a large scale study, across 101 patient cases with full PI-RADS score and histopathology, to compare the performance of correlated diffusion imaging in prostate cancer detection and localization to apparent diffusion coefficient maps, the most commonly used diffusion weighted imaging-derived imaging modality in cancer grading. Using threshold-based classification, experimental results showed that CDI achieves higher specificity at high sensitivity values of 90% and 95%, suggesting that CDI is well suited for scenarios where high sensitivity is crucial, such as cancer screening.
Introduction
Prostate cancer remains a leading cause of cancer-related mortality in men, but early detection significantly improves survival rates. Radiologists currently rely on Diffusion-Weighted Imaging (DWI), specifically Apparent Diffusion Coefficient (ADC) maps, to identify tumors. The underlying theory is that cancerous tissue has higher cell density, which restricts water molecule movement and results in lower ADC values. However, ADC maps often struggle to clearly distinguish between cancerous and healthy tissue, leading to clinical ambiguity. This study provides a large-scale validation of Correlated Diffusion Imaging (CDI) to determine if it can improve upon the current ADC standard.
Methodology and Data
The researchers utilized a dataset of 101 patient cases, representing a five-fold increase in data compared to the original pilot study for CDI. Each case included full PI-RADS scores and histopathology (biopsy results) to ensure diagnostic accuracy. Out of these cases, 39 had confirmed cancerous tumors.
A threshold-based classification approach was employed to compare the modalities. Expert radiologists labeled "tumor masks" to establish ground truth. A range of signal intensity thresholds was then applied to the images; pixels were classified as cancer if their values fell above (for CDI) or below (for ADC) specific thresholds. Performance was measured using Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC), with a specific focus on specificity at 90% and 95% sensitivity levels.
Results
The quantitative analysis revealed a massive performance gap between the two modalities. CDI achieved an AUC of 0.933, vastly outperforming ADC's AUC of 0.265. While the low ADC score confirms that cancer is inversely correlated with ADC values, the CDI score demonstrates a much stronger and more reliable positive correlation for detection.
For screening tools, high sensitivity is mandatory to avoid missing diagnoses. CDI proved highly effective in this regard, maintaining a specificity of 83.2% at 90% sensitivity and 59.7% at 95% sensitivity. In contrast, ADC's specificity dropped to 15.4% and 11.5% at those same sensitivity levels, respectively. This indicates that CDI is far superior at correctly identifying healthy tissue while still catching the vast majority of cancers.
Conclusion and Future Work
This study demonstrates that CDI provides significantly higher specificity at the high sensitivity levels required for effective medical screening compared to standard ADC maps. While the current work utilized simple threshold-based classification to compare raw signal power, future research will employ "discovery radiomics" and AI to extract more subtle quantitative patterns from CDI sequences. The ultimate goal is to integrate these findings into computer-aided clinical decision support systems to assist radiologists in making faster, more accurate diagnoses.
Summary updated