Prototyping

Overview

• model
• synthetic data generation
• training protocol
• testing
• analysis

---

Replication

synthetic data generation

$ python synthetic_medical_generator.py \
    --image-size "64,64,32" \
    --total-samples 1000 \
    --output-dir "../data/synthetic_medical_dataset_hdf5/"

training

$ python train.py with dataset_path="../data/synthetic_medical_dataset_hdf5/" training.batch_size="4"

testing

$ python test.py --weights ../train/checkpoints/best_model.pth --dataset ../data/synthetic_medical_dataset_hdf5/ --split test --output ./revised_backbone/test/

analysis

$ python analysis.py --train ../test/revised_backbone/train/detailed_metrics.json --val ../test/revised_backbone/val/detailed_metrics.json --test ../test/revised_backbone/test/detailed_metrics.json -o ./revised_backbone/

---

Analysis

1. OVERALL PERFORMANCE:

loss

metrics

TRAIN: Presence Acc: 0.9994, Dice: 0.9338, IoU: 0.8862

VALIDATION: Presence Acc: 0.9731, Dice: 0.8394, IoU: 0.7741

TEST: Presence Acc: 0.9650, Dice: 0.8396, IoU: 0.7762

2. STRUCTURE PERFORMANCE (Test Set):

Best performing structures (Dice score): liver: 0.9803 right_lung: 0.9791 left_lung: 0.9718

Worst performing structures (Dice score): L4: 0.7216 L3: 0.6904 L2: 0.6042

Present/Absent

Scenario analysis

Segmentation

Test set example visualisation

---

---

---

---