A cohort of 131 women who underwent 3T DCE-MRI between 2016 and 2021 prior to neoadjuvant chemotherapy was used for this retrospective study. The second subtracted post-contrast images were used for training, paired with raw breast delineation and precise manual segmentation of malignant lesions by an experienced radiologist.
V-Net was implemented in Python, utilising the PyTorch (version2.0.1) package. Preprocessing and postprocessing of images are carried out using the Python package SimpleITK (version 1.2.4). The training set is initially partitioned into a training subset (80%) and a validation set (20%). This partition helps determine the optimal number of epochs to prevent overfitting. Once this optimal number is identified, the entire training dataset is utilised for the fixed number of epochs.
To reduce confounding organs and computational burden, a two-stage workflow was employed. First, a V-Net model trained on raw breast segmentation restricted the area for lesion identification, producing a breast mask. Next, an overlap-and-tile method divided the area into smaller 3D patches for training while maintaining fine details. More precisely, to capture fine-grained features while preserving resolution, images were lightly downsampled to a common resolution and divided into smaller 3D patches. The training of a V-Net involved using these patches, containing both lesion and background areas, enhancing the network’s capabilities and reducing the risk of overfitting.
Final lesion segmentation merged results from overlapping patches. Parameters, including the percentage of training background patches and binarisation threshold, were tuned to select the best model. All thge process is summarized in figure 2.
To evaluate model performance, it is crucial to establish methods for comparing automatic and manual segmentations. We considered Dice score (DSC) and relative size difference (RSD) to take into account both overlap and volume.
