|
Artificial life with deep learning for segmentation of breast ultrasound images |
|---|---|
| รหัสดีโอไอ | |
| Title | Artificial life with deep learning for segmentation of breast ultrasound images |
| Creator | Suman Sharma |
| Contributor | Stanislav S. Makhanov, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | Breast cancer, Ultrasound, Medical imaging, Deep learning, Semantic segmentation, Artificial life, Tensor field analysis, มะเร็งเต้านม, อัลตราซาวนด์, การถ่ายภาพทางการแพทย์, การเรียนรู้เชิงลึก, การแบ่งส่วนเชิงความหมาย, ปัญญาประดิษฐ์, การวิเคราะห์สนามเทนเซอร์ |
| Abstract | Early and accurate segmentation of breast lesions in ultrasound (US) imaging is crucial for computer-aided diagnosis (CAD) and effective clinical decision-making. Although deep learning (DL) has significantly advanced medical image segmentation, conventional models often struggle with speckle noise, low contrast, and irregular lesion boundaries. This research proposes a novel Artificial Life Deep Learning (ALDL) framework that integrates biologically inspired agent-based modeling with deep neural architectures and tensor field (TF) analysis to enhance segmentation precision and interpretability. The methodology follows a multi-stage optimization process. First, an evaluation of DeepLabV3+ backbones identified ResNet50 as the optimal architecture for dual-input grayscale configurations. To address boundary irregularities, artificial life (ALife) agents are initialized via pathline flow derived from a vector field (VF) orthogonal to the Generalized Gradient Vector Flow (GGVF). This VF is transformed into a bi-directional TF to extract degenerate points (DPs), which serve as the basis for training a lightweight Decision-Making Network (DMN). The DMN guides agent movement along pathlines to form coherent closed contours, which are further refined using Active Contour principles and feature maps from the ResNet50 backbone. Ablation studies confirm the necessity of individual section, demonstrating that inclusion of DMN-guided agents improves the Dice score by approximately 7.0–12.0% compared to a standard ResNet50 baseline. Rigorous 10-fold cross-validation across three diverse datasets: Thammasat University Hospital (TUH), BUSI, and UDIAT, demonstrates the framework’s robustness and stability. The proposed ALDL framework achieved mean Dice scores of 94.84±1.63%, 94.16±1.62%, and 93.67±1.51% for TUH, BUSI, and UDIAT, respectively. Evaluation in contrast to state-of-the-art models demonstrates the better and consistent outcome of the ALDL model, exceeding recent models like FET-UNet by up to 11.26% in Dice score on the BUSI dataset and maintaining superior IoU values (up to 91.19±1.76%) across all benchmarks. By uniting ALife dynamics, TF topology, and DL, this research establishes a new paradigm for interpretable and structurally aware medical image segmentation. |
