Welcome! This project explores the quantification of differences in expert segmentations from CT and MRI images using deep learning. The authors are David Dashti and Filip Söderquist.
This work is based on the QUBIQ Challenge, which provides multi-expert annotated datasets for several medical imaging tasks.
All data is located on the server at /tf/tf/Project/Project_data:
- Training data:
/tf/tf/Project/Project_data/training_data_v2 - Validation data:
/tf/tf/Project/Project_data/validation_data_v2 - Test data:
/tf/tf/Project/Project_data/test
The challenge includes four datasets, each in its own folder, containing 2D slices and masks for:
- brain-growth: One-channel MRI, one task, seven expert masks per task
- brain-tumor: Four-channel MRI, three tasks, three expert masks per task
- kidney: One-channel CT, one task, three expert masks per task
- prostate: One-channel MRI, two tasks, six expert masks per task
Example: /tf/tf/Project/Project_data/training_data_v2/brain-growth contains the brain-growth data.
Scripts for each dataset are provided:
- brain-growth:
/tf/tf/Project/Brain_Growth.py - brain-tumor:
/tf/tf/Project/Brain_Tumor.py - kidney:
/tf/tf/Project/Kidney.py - prostate:
/tf/tf/Project/Prostate.py
Some datasets have multiple tasks. To run a specific task, set the task variable in the script to the desired task number (see comments in the code). This manual step is required to avoid server crashes when running multiple networks in succession.
Valid task values:
- brain-tumor: 1, 2, 3
- prostate: 1, 2
After setting the task, run the corresponding .py file to train the network and generate results.
- For each expert mask, a separate network is trained.
- Each network predicts on the test data.
- Predictions are ensembled (averaged) and compared to ensembled expert masks.
- Binarization is performed at nine thresholds: 0.1, 0.2, ..., 0.9.
- Results include the average DICE score per task and a matrix showing individual test image performance at each threshold.
- Architecture: U-Net with DICE loss was used throughout.
- Depth: Deeper networks (more convolution/deconvolution blocks) improved results.
- Dropout: Spatial dropout outperformed standard dropout, likely by encouraging the network to learn more robust features.
- Augmentation: Essential due to limited data. Used random rotation (10°), width/height shift (0.1), and horizontal flip (except for kidney).
- LSTM Layers: Adding LSTM layers in deconvolution blocks improved performance (except for brain-growth).
- Preprocessing: Intensity windowing was crucial for kidney (CT) images.
Average DICE scores (top leaderboard result in parentheses):
- brain-growth: 0.9034 (0.5548)
- brain-tumor task 1: 0.8547 (0.9169)
- brain-tumor task 2: 0.7781 (0.9224)
- brain-tumor task 3: 0.7566 (0.9682)
- kidney: 0.9181 (0.8532)
- prostate task 1: 0.9296 (0.5987)
- prostate task 2: 0.9075 (0.6302)
Total average: 0.864 (0.7778)
- The main challenge was the small number of training/validation samples.
- Test data did not include masks, so some training samples were held out for testing, further reducing training size.
- Results are not directly comparable to the leaderboard, as the test sets differ.
- The largest deviation was in brain-tumor tasks, possibly due to differences in data distribution or handling of multi-channel MRI.
- Despite these challenges, our results are strong, especially for kidney and prostate.
If you use this code or results, please cite this repository and the QUBIQ Challenge.
For questions or suggestions, please contact David Dashti or Filip Söderquist.