Mitral Valve Segmentation
Role
ML Engineer
Year
2025
Category
Computer Vision
Context
Academic Project
01.
The Problem
Manual segmentation of the mitral valve in echocardiography is prone to high inter-observer variability due to intrinsic speckle noise, low contrast artifacts, and rapid valve motion. Standard CNNs often struggle to distinguish the valve boundaries from surrounding tissue clutter.
I implemented a custom Attention U-Net architecture enhanced with Squeeze-and-Excitation (SE) blocks for channel-wise feature recalibration and Attention Gates to suppress irrelevant background regions. The data pipeline notably employed Non-local Means Denoising and CLAHE for signal enhancement, alongside extensive Elastic Deformations to simulate organic tissue variations.
02.
The Approach
03. The Result
The model achieved precise boundary delineation, optimized via a hybrid Dice-BCE loss function to balance segmentation overlap and pixel-wise classification. By merging amateur and expert datasets, the solution demonstrated strong generalization capabilities, validated through rigorous IoU tracking.
EVIDENCE_LOG
