ABSTRACT Data augmentation is crucial for training deep learning models in breast cancer detection and segmentation, but conventional methods can cause annotation errors and visual artefacts, which harm instance‐level localisation, generalisation and clinical reliability. This study aims to develop an annotation‐aware, real‐time data augmentation framework that preserves spatial and clinical integrity during geometric transformations to improve model robustness and performance. A real‐time augmentation framework dynamically recalculates bounding box coordinates and segmentation masks during cropping and rotation. Instance‐level annotation correction is performed on‐the‐fly within the training pipeline, without offline preprocessing, additional storage, or generative modelling and is evaluated on the DDSM, INbreast and BUSI datasets. Experimental results show consistent performance gains across all datasets and models, with detection metrics improving by an average of 4.2–4.4% and segmentation accuracy increasing by up to 9.3%. The real‐time implementation achieves low preprocessing latency (≈0.12 s per batch, 3.75 ms per image), enabling high‐throughput training without added computational overhead. By preserving annotation integrity during geometric transformations, the proposed framework provides a computationally efficient and easily integrable solution for breast cancer imaging, with broader applicability to other medical image analysis tasks requiring precise spatial annotations.
Mahichi et al. (Thu,) studied this question.