HipMRI Study Segmentation with an Improved 2D UNet (Task 3) | s4744546

[baulch-m]

Author

Marcus Baulch s4744546

Summary

This pull request introduces a semantic segmentation of HipMRI images using an improved U-Net architecture. The model segments anatomical structures into 6 distinct classes and achieves a Dice coefficient of 0.8777 on the test set, exceeding the minimum requirement of 0.75.

Objectives

• Implement improved U-Net architecture for prostate image segmentation
• Achieve >0.75 Dice coefficient on HipMRI dataset
• Support multi-class (6 classes) segmentation of prostate structure
• Provide training, evaluation, and visualisation capabilities

Key Features

Architecture Improvements

• Residual Connections: Added ResNet-style skip connections within encoder/decoder blocks for improved gradient flow
• Batch Normalisation: Stabilises training and improves convergence
• Multi-Scale Feature Extraction: 5-level encoder-decoder with skip connections

Training Enhancements

• Combined Loss Function: 60% Dice Loss + 40% Cross-Entropy
  • Dice Loss directly optimises the evaluation metric
  • Cross-Entropy provides stable pixel-wise gradients
• Data Augmentation: Random flips and rotations (15 degrees) applied during training

Dataset

• HipMRI Study Open Dataset: 12,660 2D MRI slices
• 6-Class Segmentation: Background, body outline, bone, bladder, rectum, prostate
• Pre-split Data: 90% train / 5% validation / 5% test

Results

Performance Metrics

Metric	Value
Test Dice Score	0.8777
Validation Dice	0.8654
Minimum Required	0.7500
Training Time	~10-15 min (A100 GPU)

Key Achievements

• Exceeds requirement by 17% (0.8777 vs 0.75 required)
• Consistent performance across train/val/test splits
• Stable training with combined loss function
• Robust to data augmentation

Implementation

Files Added

Core Implementation

• modules.py - Residual U-Net architecture with ResidualBlock and UNet classes
• dataset.py - DataSegmenter class with augmentation support
• train.py - Training pipeline with DiceLoss and ModelTrainer
• predict.py - Evaluation script with multi-class Dice computation and visualisation generator
• README.md - Thorough documentation of model and outputs

Dependencies

torch>=1.9.0
torchvision>=0.10.0
numpy>=1.19.0
nibabel>=3.2.0
matplotlib>=3.3.0
tqdm>=4.60.0
scipy>=1.5.0

[wangzhaomxy]

<This is an initial inspection, no action is required at this point.>

File Organizing:

• Please restore the original files in the root folder "/PatternAnalysis-2025/recognition/" and "/PatternAnalysis-2025/", all custom files must be put in the "/PatternAnalysis-2025/recognition/your_project_name_and_SID/" folder. Please make these corrections; otherwise, the merge request will be rejected.

Problem Solving:

• The algorithm solves the problem appropriately.
• Accuracy in testing dataset (Dice): 0.87. Meet the requirements.

Model and functions:

• It correctly uses PyTorch to construct the improved UNet 2D models and functions.
• NO data augmentation.
• Properly use the train/validation/test datasets.

Code design: Good.

Code comment and docstring:

• Good code comments
• Good function docstrings
• NO header block

Difficulty: Normal.

Additional Comments:

• Good commits.
• Great ReadMe design and comprehensive report content. However, as a report, discussion and conclusion section is expected to properly summarize your project.
• Good work. Well done!

[gayanku]

Marking

Good/OK/Fair Practice (Design/Commenting, TF/Torch Usage)
	Good design and implementation.
	Spacing and comments.
	No Header blocks.	-1
Recognition Problem
	Good solution to problem.
	Driver Script present.
	File structure present.
	Good Usage & Demo & Visualisation & Data usage.
	Module present.
	Commenting present.
	No Data leakage found.
	Difficulty : Normal. Normal. ImprovedUnet2D	-5
Commit Log
	Good Meaningful commit messages.
	Good Progressive commits.
Documentation
	Readme :Good.
	Model/technical explanation :Good.
	Description and Comments :Good.
	Markdown used and PDF submitted.
Pull Request
	Successful Pull Request (Working Algorithm Delivered on Time in Correct Branch).
	Feedback action require: Feedback marks possible +2 if the requested changes are made. Fix stucture for merge.	-2
	Request Description is good.
TOTAL		-8

Marked as per the due date and changes after which aren't necessarily allowed to contribute to grade for fairness.
Subject to approval from Shakes