Deep Edge Detection Methods for the Automatic Calculation of the Breast Contour

Area of Science:

• Medical imaging analysis
• Computer vision in healthcare
• Oncology and surgical outcomes

Background:

• Breast cancer conservative treatment (BCCT) offers comparable survival rates to mastectomy with improved cosmetic outcomes.
• Objective evaluation of BCCT aesthetic results is lacking, hindering standardized assessment.
• Current methods for automatic aesthetic evaluation rely on breast contour detection from digital photographs.

Purpose of the Study:

• To develop an improved method for automatic breast contour detection in digital photographs for BCCT aesthetic evaluation.
• To enhance the accuracy and generalization capabilities of models used for classifying BCCT cosmetic results.

Main Methods:

• Proposed a novel neural network solution to replace the conventional Sobel filter for breast contour detection.
• Utilized a shortest path algorithm on the output of the neural network for precise contour identification.
• Trained and tested models on existing and new, more variable datasets to assess generalization.

Main Results:

• Achieved state-of-the-art results on a benchmark dataset for breast contour detection.
• Demonstrated superior generalization capabilities on a new, diverse dataset compared to previous deep models.
• The neural network effectively learns representations for breast-torso wall edges, outperforming the Sobel filter.

Conclusions:

• The proposed neural network-based approach significantly improves breast contour detection for BCCT aesthetic evaluation.
• This method offers enhanced accuracy and better generalization, addressing limitations of current image processing techniques.
• The approach is simple to train and test, promoting reproducibility in objective classification of BCCT outcomes.