Project 3: Improved 2D UNet for HipMRI Prostate Segmentation (Normal Difficulty) - s4884308 by prabhjotsingh1313 · Pull Request #88 · shakes76/PatternAnalysis-2025

Initial project scaffold for HipMRI prostate segmentation

prabhjotsingh1313

committed

Add inital dataset with necessary imports and implemented build_pairs()

prabhjotsingh1313

committed

Add label discovery function to scan segmentation files and extract unique label IDs from training set

prabhjotsingh1313

committed

Implemented PyTorch dataset for paired MRI/seg nifti files and added z-score normalisation and one-hot encoding with risizing of 256x256

prabhjotsingh1313

committed

added get_data_loaders for building train/val/test splits and label discovery

prabhjotsingh1313

committed

Add dataloaders for train, validation, and test splits with batching, shuffling and metadata return

prabhjotsingh1313

committed

add center_crop function for spatial alignment in Unet

prabhjotsingh1313

committed

Add conv_block module with dual conv layers, batch norm and ReLU

prabhjotsingh1313

committed

Add Down module for 2x downsampling using max pooling with a convolutional block to reduce spatial size and increase feature depth in the improved Unet

prabhjotsingh1313

committed

Implemented Up block for Unet upsampling wih skip connections

prabhjotsingh1313

committed

Added ImprovedUNet2D with deeper layers and dilated bottleneck for segmentation

prabhjotsingh1313

committed

Implemented dice_per_channel() to compute dice score and implemented dice_loss() for model training

prabhjotsingh1313

committed

Added training and validation script for unet2D using dice loss in run_one_epoch()

prabhjotsingh1313

committed

Implemented evaluate_dice() function for per channel dice evaluation, uses no_grad() and eval mode

prabhjotsingh1313

committed

Implemented plot_training_curves to plot and save training/validation loss curves

prabhjotsingh1313

committed

Initialise README

prabhjotsingh1313

committed

Refactored improved Unet into recognition/improved2DUnet module

prabhjotsingh1313

committed

Remove duplicate scripts from recognition root after module refactor

prabhjotsingh1313

committed

Added main training script with full pipeline for improved2DUnet. Included arguement parsing, data loading, training/validation loop with dice loss, checkpoint saving and loss plotting

prabhjotsingh1313

committed

Added utility for creating fixed colourmap for multi-class segmentation masks

prabhjotsingh1313

committed

Added batch visualisation function for MRI segmentation and prostate overlay

prabhjotsingh1313

committed

Added full inference pipeline including load checkpoint, run prediction, compute dice per channel and overlay prostate mask

prabhjotsingh1313

committed

Implemented inference script with argument parsing for data, checkpoint, device, and output visualisation

prabhjotsingh1313

committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Project 3: Improved 2D UNet for HipMRI Prostate Segmentation (Normal Difficulty) - s4884308#88

Project 3: Improved 2D UNet for HipMRI Prostate Segmentation (Normal Difficulty) - s4884308#88
prabhjotsingh1313 wants to merge 33 commits intoshakes76:topic-recognitionfrom
prabhjotsingh1313:topic-recognition

Commits on Oct 29, 2025

Initial project scaffold for HipMRI prostate segmentation

Add inital dataset with necessary imports and implemented build_pairs()

Add label discovery function to scan segmentation files and extract unique label IDs from training set

Implemented PyTorch dataset for paired MRI/seg nifti files and added z-score normalisation and one-hot encoding with risizing of 256x256

added get_data_loaders for building train/val/test splits and label discovery

Add dataloaders for train, validation, and test splits with batching, shuffling and metadata return

add center_crop function for spatial alignment in Unet

Add conv_block module with dual conv layers, batch norm and ReLU

Add Down module for 2x downsampling using max pooling with a convolutional block to reduce spatial size and increase feature depth in the improved Unet

Implemented Up block for Unet upsampling wih skip connections

Added ImprovedUNet2D with deeper layers and dilated bottleneck for segmentation

Implemented dice_per_channel() to compute dice score and implemented dice_loss() for model training

Added training and validation script for unet2D using dice loss in run_one_epoch()

Implemented evaluate_dice() function for per channel dice evaluation, uses no_grad() and eval mode

Implemented plot_training_curves to plot and save training/validation loss curves

Initialise README

Refactored improved Unet into recognition/improved2DUnet module

Remove duplicate scripts from recognition root after module refactor

Added main training script with full pipeline for improved2DUnet. Included arguement parsing, data loading, training/validation loop with dice loss, checkpoint saving and loss plotting

Added utility for creating fixed colourmap for multi-class segmentation masks

Added batch visualisation function for MRI segmentation and prostate overlay

Added full inference pipeline including load checkpoint, run prediction, compute dice per channel and overlay prostate mask

Implemented inference script with argument parsing for data, checkpoint, device, and output visualisation

Commits on Oct 30, 2025

Added README documentation for Improved 2D U-Net project. Added details about algorithm description, dataset structure, preprocessing pipelines, dependencies, usage examples, quantitative results a…

Added training_curves photo

Add training curves visualization

Updated training_curves.png

Removed a note sentences

Add all visualization images (training curves, predictions, overlays)

Add prediction and overlay visualization references to README

Update README file structure section with complete visualizations

Specified the project number (Project 3) in the topic

Specify development environment (Google Colab with GPU)