Project 7 - Improved 3D U-Net for Prostate MRI Segmentation (Abhya Garg- 48299785)#285

Open

abhyagarg22 wants to merge 63 commits into

shakes76:mainfrom

abhyagarg22:topic-recognition

Commits on Sep 21, 2025

Added recognition branch for tasks
shakes76
committed

Commits on Oct 28, 2025

Commits on Oct 29, 2025

Commits on Nov 1, 2025

Commits on Nov 2, 2025

Commits on Nov 5, 2025

fix: correct over-segmentation and update predict.py for better mask visualization
abhyagarg22
committed
Merge branch 'topic-recognition' of https://github.com/abhyagarg22/PatternAnalysis-2025 into topic-recognition
abhyagarg22
committed
fix: correct label preprocessing to use prostate class (label == 5)
abhyagarg22
committed
Fix loss weighting for prostate class imbalance and enable fine-tuning
abhyagarg22
committed
Added average dice coefficient
abhyagarg22
committed
Fixed train.py and dataset.py
abhyagarg22
committed
fix: save smooth probability map instead of binary mask in predict.py
abhyagarg22
committed
fix: apply softmax-based inference for multiclass segmentation in predict.py
abhyagarg22
committed
fix: remove hard thresholding in predict.py inference
abhyagarg22
committed
fix: rescale sigmoid output for better visualization and precision
abhyagarg22
committed
fix(predict.py): ensure sigmoid applied to model output for single-channel (binary) segmentation
abhyagarg22
committed
fix(dataset): correctly match MRI and label files by patient-week ID
abhyagarg22
committed
Add validation split and fix training dataloader in train.py
abhyagarg22
committed
Improve training stability and fix underfitting (z-score normalization + weighted BCE)
abhyagarg22
committed
Added Dice + Focal combined loss to fix flat predictions and improve segmentation contrast
abhyagarg22
committed
fix: ensure sigmoid normalization before saving predictions for visualization
abhyagarg22
committed
add: multi-MRI prediction loop for UNet3D inference
abhyagarg22
committed
Updated ImprovedUNet3D to 6-class segmentation, switched to CrossEntropyLoss, fixed dataset label handling, and adjusted predict + visualize for multi-class output
abhyagarg22
committed
Refactor: multi-class support + proper affine saving
abhyagarg22
committed
Increase epoch
abhyagarg22
committed
Adjust epoch due to runtime error
abhyagarg22
committed

Commits on Nov 6, 2025