A Multi-Institutional Open-Source Benchmark Dataset for Breast Cancer Clinical Decision Support using Synthetic Correlated Diffusion Imaging Data

A new form of magnetic resonance imaging (MRI) called synthetic correlated diffusion (CDI^s) imaging was recently introduced and showed considered promise for clinical decision support for cancers such as prostate cancer when compared to current gold-standard MRI techniques such as T2-weighted (T2w) imaging, diffusion-weighted imaging (DWI), and dynamic contrast-enhanced (DCE) imaging  [1]. However, the efficacy for CDI^s for other forms of cancer such as breast cancer has not been as well-explored nor have CDI^s data been previously made publicly available. The development of computer-aided clinical decision support for breast cancer using CDI^s has begun to be analyzed and shown to have superior results compared to other gold-standard imaging for the prediction of breast cancer patient response from neoadjuvant chemotherapy prior to treatment  [2]. Motivated to advance efforts in the development of computer-aided clinical decision support for breast cancer using CDI^s for diagnosis, prognosis/grading, treatment planning and more, we introduce Cancer-Net BCa, a multi-institutional open-source benchmark dataset of volumetric CDI^s imaging data of breast cancer patients with detailed annotation metadata for each patient. We further examine the demographic and grade diversity of the Cancer-Net BCa dataset to gain deeper insights into potential biases. The Cancer-Net BCa benchmark dataset has been made publicly available ¹¹1https://www.kaggle.com/datasets/amytai/cancernet-bca as a part of a global open-source initiative dedicated to accelerating advancement in machine learning to aid clinicians in the fight against cancer.

To construct the Cancer-Net BCa benchmark dataset, we produced CDI^s acquisitions for a pre-treatment (T0) patient cohort of 253 patient cases across 10 institutions via the American College of Radiology Imaging Network (ACRIN) 6698/I-SPY2 study [3, 4, 5, 6]. More specifically, acquisitions were conducted with a four b-value imaging protocol (0 s/mm², 100 s/mm², 600 s/mm², 800 s/mm², 3-direction) on a 1.5 or 3.0 Tesla scanner using a dedicated breast radiofrequency coil. The pixel spacing for the acquisitions ranged from 0.83 mm to 2.08 mm with a median of 1.29 mm, with both slice thickness and spacing between slices ranged from 4.0 to 5.0 mm with a median of 4.0. The native and synthetic signals produced via a signal synthesizer were mixed together to obtain a final CDI^s signal [1]. Each patient case is also associated with one of three possible SBR grades: I (Low), II (Intermediate), and III (High). Example images from each SBR type is shown in Fig. 1. The pCR state after neoadjuvant chemotherapy (No pCR/pCR) is also provided for each patient, with an example of each pCR state shown in Fig. 2.

The demographics of the Cancer-Net BCa dataset is shown in Table 1. It can be seen that the White race dominates the data, comprising of 70.8% of the patients in the dataset, illustrating a severe race bias towards White patients. Additionally, Fig. 3 (top), it can be seen that the majority of the patients are between 30 to 70 years old (95.7%), indicating that very young patients ($~{}\leq$ 29) and very old patients ($~{}\geq$ 70) could be underrepresented in the dataset. On the other hand, the genetic subtype in the dataset is more fairly distributed with each subtype represented in at least 10% of the patients whereas the lesion type is more biased towards multiple masses and single mass as seen in Fig. 4 upper left and right respectively. In addition, the longest diameter on the MRI (MRLD) is also biased towards the range of 2 to 4 cm with less representation from patients in the other diameter ranges as seen in Fig. 3 (bottom).

The grade distribution and pCR division are shown in bottom half of Fig. 4, indicating an uneven distribution in SBR grade, significantly skewed towards Grade III (High) and shows that more patients with no pCR (67.6%) compared to those who achieved pCR after neoadjuvant chemotherapy (32.4%). Noting the demographic, grade, and pCR imbalances, it is recommended to use algorithms and strategies that account for the imbalanced dataset such as data sampling, re-balancing of the classes, and balanced loss functions. Furthermore, these imbalances should be considered when evaluating systems developed on this dataset such as with balanced metrics such as per-class precision and recall.