AI Inference trace

[vikashg]

Is your feature request related to a problem? Please describe. What we want is that, along with model inference, there is a file (either JSON, TXT, or XML) which contains all the information required to trace back the result to the source.

Often in hospital environments, the doctors and AI scientists want to know what operations are performed on the patient data. They should be able to look it up when needed. This ability will instill confidence in making decisions on patients as well as do root cause analysis on the data.

The goal of this feature request is to decide on a schema about what are the important datapoints that should be collected for inference monitoring.

Here is a proposed schema/list of attributes to be collected during inference.

• Image Series Description
• Image intensity characteristics
• Whether it was converted to NIfTI or NumPy before absorbing into MONAI
• Which reader was used for reading the images (ITK, NiBabel, or something else)
• List of transforms in order
• Operations performed on the output (sigmoid, thresholding values)
• Precision of the GPU and model used
• Model Version

[vikashg]

I am also not sure if this should be in MONAI Deploy App SDK or MONAI Core. I believe it should be a feature in the MONAI Core. The tooling might need to change for monai deploy.