How damidirad/ai4mi_project works
This repository is a SegTHOR medical-image segmentation project. Its user workflow is to prepare 3D CT data as 2D slices or tiles, train a segmentation network, then inspect predictions and metrics; an optional tiled route evaluates reconstructed volumes in original geometry. The graph preserves both the documented baseline features and the tiled extensions visible in the tree. Some implementation links are dashed because the relevant call sites were not included in the excerpts.
Main entry points: main.py Baseline training · slice_segthor.py Slice preprocessing · train_tiled.py Tiled training · tiled_dataset.py Tiled dataset
Components 22
Data preparation
- Slice preprocessing
slice_segthor.py - Synthetic data
gen_two_circles.py - Voxel geometry
geometry.py - Preprocessing metadata
preprocessing_metadata.py - Tile layout and writing
tiling.py - Slice dataset
dataset.py - Tiled dataset
tiled_dataset.py
Segmentation models
- ENet
ENet.py - 2.5D ENet
ENet_2_5d.py - ShallowNet
ShallowNet.py
Training and evaluation
- Baseline training
main.py - Segmentation losses
losses.py - Tiled training
train_tiled.py - Tiled losses
tiled_losses.py - Volume evaluation
evaluate_tiled.py
Results and geometry
- Tile reconstruction
reconstruct_tiled.py - Slice stitching
stitch.py - Segmentation viewer
- Metric plotting
plot_metrics.py - Metrics and utilities
utils.py
Other components
- Researcher
- CT volumes
Connections 25
- Researcher → Slice preprocessing: prepares data
readme.md - CT volumes → Slice preprocessing: reads volumes
- Researcher → Synthetic data: generates toy data
readme.md - Slice preprocessing → Voxel geometry: uses geometry
slice_segthor.py - Slice preprocessing → Preprocessing metadata: records geometry
slice_segthor.py - Slice preprocessing → Tile layout and writing: writes tiles
slice_segthor.py - Researcher → Baseline training: trains model
readme.md - Baseline training → Slice dataset: loads slices
main.py - Baseline training → ENet: selects network
main.py - Baseline training → 2.5D ENet: selects network
main.py - Baseline training → ShallowNet: selects network
main.py - Baseline training → Segmentation losses: uses loss
main.py - Baseline training → Tiled training: dispatches tiled
main.py - Tiled training → Tiled dataset: builds datasets
- Tiled training → Tiled losses: builds losses
train_tiled.py - Tiled training → Volume evaluation: validates volumes
train_tiled.py - Tiled dataset → Preprocessing metadata: loads records
tiled_dataset.py - Tiled dataset → Tile layout and writing: uses layout
tiled_dataset.py - Volume evaluation → Voxel geometry: restores grid
evaluate_tiled.py - Volume evaluation → Preprocessing metadata: validates geometry
evaluate_tiled.py - Researcher → Tile reconstruction: reconstructs tiles
- Researcher → Slice stitching: stitches slices
readme.md - Researcher → Segmentation viewer: compares predictions
readme.md - Researcher → Metric plotting: plots metrics
readme.md - Baseline training → Metrics and utilities: uses metrics
main.py