How jimmycai91/dlt works
Project type and purpose
• This is an open-source lesion-tracking benchmark and toolkit.
• Primary tasks:
• Convert raw DeepLesion PNGs to NIfTI volumes (DL_save_nifti.py)
• Provide lesion pair annotations (JSON files in data/)
• Run evaluation code (evaluation.py) on predictions.
File structure
• Top-level scripts:
• DL_save_nifti.py – conversion from PNG to .nii.gz
• evaluation.py – scoring/prediction evaluation
• Data directory:
• train.json, valid.json, test.json – lesion_pair_annotation records
• DLTMix.json – combined annotation set
• Config/build:
• No heavy build system; it’s invoked via simple CLI scripts.
README for insights
• Describes dataset format and fields in lesion_pair_annotation
• Points to external DeepLesion source (NIH) as upstream data
• No existing service or web UI—purely batch scripts & data.
Main components
a. External DeepLesion Data Source
– Raw PNG slices (external repository)
b. Data Conversion Module
– DL_save_nifti.py
– Input: DeepLesion PNGs → Output: NIfTI subvolumes
c. Annotation Store
– JSON files under data/ (train/valid/test)
d. Deep Lesion Tracker (DLT) Model
– *Not included in this repo*, but implied: training & inference code
e. Evaluation Module
– evaluation.py reads model outputs + JSON → quantitative metrics
Relationships & data flows
• Annotation JSONs reference those NIfTI files (bidirectional link)
• evaluation.py consumes predicted outputs + ground truth JSON → evaluation metrics
Architectural patterns & design principles
• Pipeline pattern: clear, linear transform steps (convert → annotate → infer → evaluate)
• Separation of concerns:
• Conversion (data ingestion)
• Annotation (data labelling)
• Model (algorithmic core)
• Evaluation (metrics)
• Extensibility point: swapping in a new tracker model or evaluation metric is easy by following JSON I/O spec.
Technologies & formats
• Python scripts (DL_save_nifti.py, evaluation.py)
• NIfTI (.nii.gz) volumes
• JSON annotation schema
• CLI invocation (bash commands)
• Labels on every component (e.g., “DL_save_nifti.py”)
[External DeepLesion PNG Repository]
|
| (.png slices)
v
[DL_save_nifti.py] ––> [NIfTI Volume Store]
|
| (references)
v
[Annotation Store (train/valid/test JSON)]
|
| (consumed by)
v
[DLT Model (Training/Inference)]
|
| (predictions JSON)
v
[evaluation.py] ––> [Evaluation Metrics Output]