Semantic Segmentation (Medical Imaging)

Overview

Developed a semantic segmentation model for ocular diagnosis (LG staining) images to identify affected regions.

Responsibilities

• Prepared and cleaned medical image datasets
• Designed and trained UNet-based architecture
• Evaluated segmentation performance using domain-specific metrics

Approach

• 2D UNet architecture
• Data augmentation for limited datasets
• Pixel-wise classification

Model Architecture

Training Pipeline

flowchart LR
    A[Raw Medical Images] --> B[Preprocessing]
    B --> C[Data Augmentation]
    C --> D[UNet Training]
    D --> E[Validation]
    E --> F[Model Output]

Inference Pipeline

flowchart LR
    A[Input Image] --> B[Preprocessing]
    B --> C[Model Inference]
    C --> D[Segmentation Mask]

Tech

TensorFlow · Keras · UNet

Impact

• Enabled automated analysis of ocular conditions
• Reduced manual annotation workload
• Improved diagnostic consistency

Sample Results

Raw	Segmented