DRAW Single Machine Workflow

DRAW can be used a Desktop App as well!

Architecture

DICOM Files Placement:
- Place DICOM files in the designated input folder. This can be automatically done from Scanning Service or manually placed.
Automatic Processing:
- The system automatically detects new files in the input folder.
- Sends the detected DICOM files to the NNUNet model.
NNUNet Prediction:
- The NNUNet model analyzes the received images.
- Generates predictions for each relevant area within the images.
Combine Overlapping Predictions:
- Predictions from areas covered by multiple images are combined or merged.
- Creates a single, unified prediction for those areas.
Prediction Export:
- Combined predictions are exported to the specified output folder in the designated format.

Change the value of DICOM_WATCH_DIR in draw/config.py and run by any of the following in the command line:

python main.py start-pipeline #Preferred way

or

python run.py #Will be deprecated in the future

NNUNet Model: Trained model for prediction
DICOM Files: Standard format for storing medical images, such as MRI scans or CT scans. Files come from scanning service
Input Folder: Contains DICOM files to be processed
Output Folder: Stores exported predictions.
SQL Database: Tracks the status of predictions. Acts as a Queue.

Enables multiple predictions to be executed simultaneously by multiprocessing. 2 models will be run in parallel
Batch size of processing is 1
Specifications of Prediction Server:
- Intel® i7
- Nvidia® Quadro T1000 8GB
- 32 GB RAM

In this low-traffic scenario, a database is a convenient choice for both queuing and analytics.

Streamlines management: It keeps everything in one place, making it easier to track prediction status and results.
Direct access for analytics: You can easily access prediction data for analysis, helping you understand trends and improve the process.
Efficiency for low traffic: It's well-suited for environments with lower prediction volumes.
Doubles as storage: It serves as both a queue for managing tasks and a storage system for prediction data.

This two-in-one approach simplifies the overall architecture and promotes efficient data analysis.