JSON configuration for deep learning based segmentation

JSON-format configuration files are used to specify the input data format, label maps to be exported, and various pre-processing options (including cropping, resizing, and populating image channels). Processed data may be exported to NIfTI (3D volume), NRRD (3D volume) or HDF5 (2-D slice-wise fashion or 3-D volume) format, as required by the user.

Sample configuration

JSON outline

Sample settings (Click to expand)

Training	Inference
Templates: basic and advanced	View configurations for pre-trained models distributed with CERR.

Customizing settings

I/O

• Supported input types: scan, structure
• Supported output types: labelMap, DVF

Input processing settings

(Required)

1. scan : This field specifies scan identifiers and pre-processing parameters. Settings for multiple scans can be input as nested lists.
See supported settings

For training only

2. inputFileType : May be "DICOM" or "CERR".

3. dataSplit : This setting is used to randomly split data into training, validation, and testing (inference) sets per user-specified proportions.
Default: testing split of 100% is assumed.
"dataSplit": [0 0]

(Optional)

1. structure: Specify names of structures to be exported and associated numeric label.

structure": {"name": "GTV_copy"}

Data export settings for AI model

Model I/O related parameters.

1. modelInputFormat, modelOutputFormat : Data formats input to & output by the segmentation model. Supported options: "NRRD","NIFTI","H5"(default).

2. exportedFilePrefix : Prefix for exported H5 file names. Default: "inputFileName".

3. passedScanDim : May be "2D" (one H5 file per scan slice) or "3D" (one NIfTI/NRRD/H5 file per scan volume).
   Note: (1) passedScanDim is assumed to be "3D" for NIfTI or NRRD model I/O formats.
   (2) If multiple channels are specified, passedScanDim must be "2D". This produces one 3D (multi-channel) H5 file per slice.

Output data processing settings

Optional fields for inference only:

1. strNameToLabelMap: May be either (a) dictionary of output structure names and corresponding numeric labels returned by the segmentation model. Required dictionary fields are: "structureName" and "value". or (b) Name of JSON file containing the structure name-to-label dictionary, generated by the segmentation wrapper.
   

   Syntax:
   (a) "strNameToLabelMap":[{"structureName" : "structName1", "value" : 1}, {"structureName" : "structName2", "value" : 2}]"
   (b) "strNameToLabelMap": "labelMapFileName.json"

2. postProc: List of parameter dictionaries (one per structure) for post-processing segmentations returned by DL models.
   

   Syntax:
   "postProc" : {"Struct1": [{"method": "method1", params" : {"param1": val1}}],
   "Struct2": [{"method": "method2", params" : {"param2": val2}}]}

   Supported methods and associated parameters are listed in the table below.

Method	Parameters	Description	Usage
getLargestConnComps	numCC	Returns 'numCC' largest connected components in structure mask.	"numCC" : n (numeric)
getLargestOverlappingComp	roiName	Returns largest connected component that overlaps with selected ROI.	"roiName" : "str1"
getSegInROI	roiName	Returns intersection of auto-segmented structure with selected ROI.	"roiName" : "str1"
removeBackgroundFP	threshold	Removes false positive detections outside the patient's outline. "threshold" (HU) is the intensity used to separate background (air) voxels.	"threshold ": -400

Additionally, custom post-processing can be applied by specifying the corresponding function name and parameters through the "method" and "params" fields, respectively. Custom functions should follow the signature [outputMask3M, planC] = func_name(strNum,paramS,planC).

Example:
"postProc" : {"LungNodule": [{"method": "getLargestConnComps", "params" : {"numCC": 1}}]}

2. inputStrNameToLabelMap : For models requiring input label maps, structure names, associated scan identifiers and labels are specified.

   Syntax:
   "inputStrNameToLabelMap" :[{ "structureName" : "Str1", "value" : 1, "assocScanIdentifier": { "scan_identifier" : "value" }},
   {"structureName": "Str2","value": 2, "assocScanIdentifier":{"scan_identifier":"value"}}]

Sample JSONs

• basic settings for building segmentation models
• advanced settings for building segmentation models
• sample settings for deploying segmentation models.

---

Pre-trained model Library

Configuration files for pre-trained models.