Hybrid MICO-LAC Segmentation with Panoptic Tumor Instance Analysis for Dense Breast Mammograms.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Dense breast tissue presents challenges in mammography due to inhomogeneity, low contrast, and complex tumor morphology.
• Accurate segmentation of tumors in mammograms is crucial for diagnosis and treatment planning.
• Existing methods often struggle with precise boundary delineation and distinguishing individual tumor instances.

Purpose of the Study:

• To develop a clinically driven hybrid segmentation framework for analyzing dense breast tissue in mammographic images.
• To improve the accuracy and interpretability of tumor segmentation, including instance segmentation for structural analysis.
• To address limitations of existing methods in handling challenging imaging conditions and complex tumor features.

Main Methods:

• Integration of Multiplicative Intrinsic Component Optimization (MICO_2D) for bias field correction.
• Application of a distance-regularized multiphase Vese-Chan level-set model for initial segmentation.
• Localized refinement using Localized Active Contours (LAC) with Local Image Fitting (LIF) energy and Gaussian regularization.
• Incorporation of a panoptic-style tumor instance segmentation stage for decomposing connected regions.
• Evaluation on MIAS and INBreast datasets using Cranio-Caudal (CC) and Medio-Lateral Oblique (MLO) views.

Main Results:

• The hybrid framework demonstrated competitive performance against U-Net and other deep learning architectures.
• Achieved strong spatial agreement with reference segmentations, evidenced by Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) metrics.
• Showcased robust performance under various perturbations (noise, blur, rotation) and clear separability between cancerous and non-cancerous tissues in feature space visualizations (t-SNE, UMAP).
• Enabled detailed structural analysis of tumor multiplicity and spatial organization, enhancing interpretability.

Conclusions:

• The proposed hybrid panoptic framework is effective and robust for comprehensive dense breast tumor analysis in mammography.
• The framework offers improved interpretability and clinical relevance compared to conventional segmentation methods.
• Results emphasize the importance of reproducibility and conservative statistical assessment in medical image analysis research.