Presegmenter Cascaded Framework for Mammogram Mass Segmentation.

Accurate segmentation of breast masses in mammogram images is essential for early cancer diagnosis and treatment planning. Several deep learning (DL) models have been proposed for whole mammogram segmentation and mass patch/crop segmentation. However, current DL models for breast mammogram mass segmentation face several limitations, including false positives (FPs), false negatives (FNs), and challenges with the end-to-end approach. This paper presents a novel two-stage end-to-end cascaded breast mass segmentation framework that incorporates a saliency map of potential mass regions to guide the DL models for breast mass segmentation. The first-stage segmentation model of the cascade framework is used to generate a saliency map to establish a coarse region of interest (ROI), effectively narrowing the focus to probable mass regions. The proposed presegmenter attention (PSA) blocks are introduced in the second-stage segmentation model to enable dynamic adaptation to the most informative regions within the mammogram images based on the generated saliency map. Comparative analysis of the Attention U-net model with and without the cascade framework is provided in terms of dice scores, precision, recall, FP rates (FPRs), and FN outcomes. Experimental results consistently demonstrate enhanced breast mass segmentation performance by the proposed cascade framework across all three datasets: INbreast, CSAW-S, and DMID. The cascade framework shows superior segmentation performance by improving the dice score by about 6% for the INbreast dataset, 3% for the CSAW-S dataset, and 2% for the DMID dataset. Similarly, the FN outcomes were reduced by 10% for the INbreast dataset, 19% for the CSAW-S dataset, and 4% for the DMID dataset. Moreover, the proposed cascade framework's performance is validated with varying state-of-the-art segmentation models such as DeepLabV3+ and Swin transformer U-net. The presegmenter cascade framework has the potential to improve segmentation performance and mitigate FNs when integrated with any medical image segmentation framework, irrespective of the choice of the model.