AI-driven diffusion weighted imaging-based non-contrast protocol for breast cancer diagnosis: a multicentre, multidimensional validation study.

Background: Contrast-enhanced breast MRI, while highly sensitive, faces limitations including complexity, long acquisition times, and reliance on gadolinium-based contrast agents. Noncontrast diffusion-weighted imaging (DWI) offers an ultrafast alternative, but its diagnostic accuracy has been insufficient for standalone use. We investigated whether a deep learning (DL) model could enable accurate breast cancer diagnosis using only DWI. Methods: This four-centre study included 2493 patients with pathologically confirmed breast lesions. A DWI-based model (DWI-DL) was developed using data from 1286 patients at Peking University People's Hospital (PKUPH; January 2015 to July 2021), who were randomly divided into three cohorts at a ratio of 6:2:2 (train n = 774, validation n = 256, test n = 256). Independent testing used three external cohorts (n = 661) and a prospective cohort from PKUPH that was retrospectively analysed (n = 546; August 2021 to September 2022). The diagnostic performance of DWI-DL model was compared to that of an abbreviated enhanced-based (AE-DL) model and two expert radiologists. Subgroup analysis assessed performance across external and prospective cohorts. A multireader multicase validation, using DW-DL model and selective contrast-enhanced workflow, was conducted by 12 radiologists from ten institutions to evaluate the clinical utility of DWI-DL in assisting diagnosis. Performance was assessed using the area under the curve (AUC) of receiver operating characteristics. The multireader multicase study was registered with the China Clinical Trial Registry (ChiCTR2500095953). Findings: DWI-DL model demonstrated diagnostic performance across all cohorts that was comparable to the AE-DL model (AUC: 0.771-0.912 vs. 0.780-0.898, p = 0.36-0.98). DWI-DL outperformed two expert radiologists interpreting DWI alone (AUC: 0.781-0.858 vs. 0.714-0.770, p range <0.0001-0.023). In the multireader multicase validation, the AI-guided selective sequence protocol was non-inferior to the full protocol (Protocol E vs. C, AUC: 0.834 [95% CI: 0.785, 0.883] vs. 0.835 [95% CI: 0.789, 0.881], difference: -0.001 [95% CI: -0.029, 0.027], p = 0.94), while simultaneously reduced interpretation time by 55.5% (Protocol E vs. C, mean time: 59.6 [43.5] vs. 134.0 [72.4] seconds, p < 0.0001). Interpretation: A DL model using only non-contrast DWI can accurately diagnose breast cancer. The DWI-DL model demonstrated robust performance, comparable to the AE-DL model and surpassing human experts in DWI interpretation, while reducing interpretation time. However, its performance was lower than that of expert radiologists interpreting standard breast MRI. By improving diagnostic efficiency without affecting accuracy, our workflow cooperating the DWI-DL model and selective contrast-enhanced sequence presents a promising, rapid, and safe tool with the potential to streamline the clinical diagnosis of breast cancer. Funding: This work was supported by the National Natural Science Foundation of China (82471964) and Peking University People's Hospital Scientific Research Development Funds (RDGS2022-10).