Our results shows that breasts with higher density (ACR C-D) presented greater reductions in parenchymal density after neoadjuvant chemotherapy compared to predominantly fatty breasts (ACR A-B). Qualitative evaluations adhered to the BI-RADS lexicon, while quantitative density was determined through AI software. This is consistent with the evidence that high mammographic density (ACR C-D) is a negative prognostic factor and it could be considered a predictor of treatment response, similar to the absence of hormone receptors or the presence of Ki67 > 20%. We also found that younger women and those in the perimenopausal stage exhibited the most notable alterations in breast density after undergoing neoadjuvant chemotherapy. Our results suggest that changes in breast density can be quantitatively assessed and may serve as viable imaging biomarkers for monitoring therapeutic response in breast cancer.

The moderate agreement among radiologists and the substantial agreement with AI tools underline the potential for integrating AI in routine assessments to improve the accuracy and reproducibility of breast density evaluations. AI applied to imaging may also predict tumor response before neoadjuvant chemotherapy begins. Our study also utilized a commercial AI tool, known for its robust volumetric breast density assessment, that uses machine learning algorithms to perform automated measurements, providing an independent assessment that can be compared with the in-house developed AI tool and radiologists' evaluations, ensuring consistency across different assessment methods. Table 2 shows the different levels of agreement among radiologists and between radiologists and AI tools, underscoring the challenges and potential of integrating AI into clinical practice for breast density evaluation. These findings highlight the significant variability that can occur with less experienced or untrained evaluators: the lower ICC values in these groups suggest that training and experience are crucial for reliable breast density assessment. Furthermore, the substantial agreement between the expert radiologist and both AI tools also supports the validity and reliability of these AI systems in clinical practice.