Abstract
This paper introduces Cancer-Net BCa, the first multi-institutional, open-source benchmark dataset of volumetric Synthetic Correlated Diffusion Imaging ($CDI^s$) for breast cancer. Comprising data from 253 patients across ten institutions, the dataset is designed to accelerate the development of AI-driven clinical decision support tools for predicting treatment response and tumor grading. The authors provide a detailed characterization of the dataset, including demographic and clinical metadata, while highlighting critical imbalances that researchers must address to build robust, generalizable models.
Introduction
Magnetic Resonance Imaging (MRI) is a fundamental tool in breast cancer diagnosis. However, standard techniques such as $T_2$-weighted or Diffusion-Weighted Imaging (DWI) often lack the precision required for advanced clinical decision support. Synthetic Correlated Diffusion Imaging ($CDI^s$) is an emerging modality that has demonstrated significant potential in prostate cancer by improving tumor delineation. Despite its promise, its application in breast cancer has been hindered by a lack of high-quality, multi-institutional benchmark datasets.
The Cancer-Net BCa dataset addresses this gap, providing a standardized resource for training and evaluating machine learning models on tasks such as predicting pathologic complete response (pCR) to neoadjuvant chemotherapy.
Methods and Data Modality
The dataset utilizes $CDI^s$ signal synthesis, a technique that mixes native and synthetic signals to enhance correlated diffusion signals indicative of cancerous tissue. This process is designed to better highlight the structural changes associated with malignancy compared to traditional DWI.
The imaging protocol involved a four b-value sequence ($0, 100, 600, 800\text{ s/mm}^2$) performed on 1.5T and 3.0T scanners. This imaging data is integrated with high-level clinical metadata, including genetic subtypes and pCR status—the gold standard for determining if a treatment was successful.
Dataset Description
Derived from the ACRIN 6698/I-SPY2 study, Cancer-Net BCa includes:
- Scale: 253 patient cases from 10 diverse institutions.
- Metadata: Lesion type, genetic subtype, longest tumor diameter (MRLD), Scarff-Bloom-Richardson (SBR) grade, and post-treatment pCR status.
- Accessibility: The dataset is publicly hosted on Kaggle as part of a global open-source initiative to foster collaborative research.
By incorporating data from multiple centers, the dataset captures real-world clinical variations, such as differences in hardware and operator protocols, which is essential for developing generalizable AI models.
Dataset Characteristics and Results
A primary contribution of this work is a transparent "bias audit" of the dataset. The authors analyzed the demographic and clinical distributions to inform researchers of potential pitfalls in model training.
Key findings from the dataset characterization include:
- Demographics: The cohort is predominantly White (70.8%), with lower representation of Black (10.7%) and Asian (6.3%) patients. Very young (≤ 29) and very old (≥ 70) patients are also underrepresented.
- Clinical Grade: 69.1% of cases are classified as Grade III (High), indicating a skew toward aggressive cancers.
- Treatment Response: 67.6% of patients did not achieve pCR, presenting a significant class imbalance for predictive modeling.
- Tumor Size: The majority of tumors have a diameter (MRLD) between 2 and 4 cm.
Discussion and Future Directions
The identified imbalances in age, ethnicity, and clinical grade pose challenges for AI development. Models trained naively on this data may exhibit biased performance across different populations or cancer stages.
To mitigate these issues, the authors recommend several strategies for researchers using the benchmark:
- Data Resampling: Implementing over-sampling for minority classes or under-sampling for majority classes.
- Algorithmic Adjustments: Utilizing balanced loss functions to penalize errors on underrepresented groups more heavily during training.
- Robust Evaluation: Prioritizing metrics like per-class precision, recall, and F1-score over aggregate accuracy to ensure the model isn't simply predicting the majority class.
The release of Cancer-Net BCa represents a significant step toward open-source, multi-institutional benchmarks in medical imaging, providing the community with the tools needed to advance AI-driven breast cancer care.
Summary updated