Abstract
Prostate cancer (PCa) is a leading cause of cancer-related deaths, yet clinical decision-making remains challenging due to the limitations of standard imaging. We introduce Cancer-Net PCa-Data, the first open-source benchmark dataset featuring synthetic correlated diffusion ($CDI^s$) imaging. This dataset provides 200 fully annotated patient cases, designed to accelerate the development of AI-driven clinical decision support tools. By leveraging the superior contrast of $CDI^s$, researchers can develop models that better distinguish between clinically significant and insignificant tumors compared to traditional MRI modalities.
Introduction
The current gold standard for prostate cancer diagnosis involves Magnetic Resonance Imaging (MRI) techniques, specifically T2-weighted (T2w) and Diffusion-Weighted Imaging (DWI). However, these modalities often struggle to clearly delineate "clinically significant" cancerous tissue from benign or insignificant cases. This diagnostic ambiguity can lead to over-treatment or missed aggressive cancers.
Synthetic correlated diffusion ($CDI^s$) imaging has emerged as a promising modality to address these challenges. By providing better contrast and clearer boundaries for aggressive tumors, $CDI^s$ serves as a potentially more powerful biomarker. Despite its promise, research has been hindered by the lack of publicly available $CDI^s$ datasets. Cancer-Net PCa-Data fills this critical gap, providing a standardized foundation for the global research community.
Synthetic Correlated Diffusion Imaging ($CDI^s$)
Unlike standard DWI, which measures signal attenuation at specific b-values (sensitivities to Brownian motion), $CDI^s$ captures the joint correlation of signal attenuation across multiple sensitivities. It integrates native and synthetic diffusion acquisitions in a calibrated manner to quantify the distribution of water molecules more precisely.
In this dataset, $CDI^s$ data was computed from DWI acquisitions across nine b-values ranging from 0 to 5000 $s/mm^2$. This approach allows the probing of tissue characteristics that are difficult to capture with hardware-limited native scans alone, resulting in stronger delineation of clinically significant cancerous tissue.
The Cancer-Net PCa-Data Dataset
The Cancer-Net PCa-Data dataset is derived from a cohort of 200 patient cases originally acquired at Radboud University Medical Centre as part of the PROSTATEx Challenges. The dataset includes volumetric $CDI^s$ images, prostate gland masks, tumor masks, and clinical labels such as Gleason scores and clinical significance.
To ensure high-quality ground truth, annotations were performed by radiology residents and reviewed by board-certified radiologists with over 20 years of experience. The dataset contains a total of 299 tumors, with 76 (25.4%) classified as clinically significant (csPCa) and 223 (74.6%) as clinically insignificant (insPCa).
Discussion and Limitations
While Cancer-Net PCa-Data provides a robust benchmark, several limitations must be considered. First, Gleason Grade Groups are unavailable for 62.5% of the tumors. This reflects clinical reality, as "insignificant" tumors (PI-RADS 1 or 2) are often not biopsied, leading to a lack of histological ground truth for these cases.
Second, the dataset exhibits demographic and class imbalances. Approximately 56% of patients are aged 60-69, with a lack of data for patients under 50. Furthermore, the 3:1 ratio of insignificant to significant tumors requires researchers to employ specific strategies, such as balanced loss functions or per-class metrics, to avoid skewed results. Transparency regarding these biases is vital for developing generalizable and fair AI models.
Conclusion
Cancer-Net PCa-Data represents a significant step forward in prostate cancer research. By providing the first public $CDI^s$ dataset, the authors enable the community to build and benchmark machine learning models on a modality that offers superior tissue delineation. The dataset is hosted publicly on Kaggle to facilitate widespread adoption and collaborative innovation in clinical decision support.
Summary updated