Instance Segmentation with Mask R-CNN

Scientific Image Analysis using PyTorch

Project Overview

This project implements an end-to-end instance segmentation pipeline using Mask R-CNN in PyTorch.

The objective is pixel-level localization of structured regions within complex image data.
The workflow demonstrates practical experience in scientific image processing, transfer learning, and reproducible ML experimentation.

Although applied here to MRI data, the pipeline design is adaptable to other domains such as:

• Satellite imagery
• Radar data
• Environmental monitoring
• Geospatial image analysis

---

Core Objective

Design a robust instance segmentation system capable of:

• Detecting regions of interest
• Generating pixel-level masks
• Adapting pretrained deep learning models to domain-specific image distributions

This mirrors challenges commonly found in large-scale scientific and environmental datasets.

---

Model Architecture

• Framework: PyTorch
• Model: Mask R-CNN
• Backbone: ResNet-FPN
• Transfer learning from COCO pretrained weights
• Custom adaptation of classification and mask heads

The architecture supports multi-scale feature extraction and precise spatial localization.

---

Data Pipeline

The project includes a structured data workflow:

1. Image preprocessing (resizing, normalization)
2. Mask generation and annotation formatting
3. Train/validation split
4. Batch loading using PyTorch Dataset & DataLoader
5. Reproducible experimentation setup

Pipeline designed for scalability and modular reuse.

---

Training Strategy

• Fine-tuning pretrained backbone
• Hyperparameter tuning
• Monitoring training stability
• GPU-compatible implementation
• Structured experiment reproducibility

---

Evaluation

Performance assessed using:

• Intersection over Union (IoU)
• Qualitative mask overlay visualization
• Loss monitoring across epochs

Focus on both quantitative and visual validation.

---

Technical Stack

• Python
• PyTorch / Torchvision
• NumPy
• Matplotlib
• Jupyter Notebook
• Linux-based execution environment

---

Engineering Highlights

• End-to-end deep learning workflow
• Domain adaptation using transfer learning
• Structured dataset handling
• Modular design for extension to other image domains
• Reproducible scientific ML experimentation

---

Potential Extensions

The pipeline can be extended to:

• Satellite cloud segmentation
• Weather radar echo segmentation
• Precipitation region detection
• Multi-class environmental feature extraction

---

Author

Hakimeh Khojasteh
Machine Learning Engineer | Scientific ML