Initial commit request for the workloads document. Please review and comment.

Author: Danulf

dev-experience-mock-up/documents/workloads.md

@@ -0,0 +1,127 @@
+# MONAI Deploy Workloads
+
+A number of workloads could be addressed by the MONAI deploy SDK and
+framework. This document outlines how we will categorize a
+workload. How to think about it from a deployment point-of-view. We
+will use these definitions to outline the requirements for deployment
+infrastructure.
+
+## Standard Language
+This document SHALL follow the
+guidance of rfc 2119 for terminology.
+
+## Definitions:
+
+- **Job:** when you request the computing resources to accomplish a specific
+task, that is called a “job”.
+
+- **Synchronous jobs:** this is a job that MUST run immediately. It MAY
+interact with humans or peripherals as needed.  Asynchronous jobs:
+these are jobs SHALL be scheduled to be run when resources are
+available.
+
+- **Timeout:** asynchronous jobs MAY be submitted with a timeout specified
+such that they MUST be scheduled within the timeout or they will be
+marked as failed. This is another way of representing a required
+service-level-agreement (SLA) necessary for the workload.
+
+- **Throughput:** This is the measure of the number of jobs that can be
+processed by specified hardware/software configuration over a specific
+period of time.
+
+- **Deterministic service requirements:** this is where the
+containers and services required to complete the job are known before
+the job is scheduled.
+
+- **Cyclic service requirements:** this is where, to
+complete the job, the container or containers may be executed in a
+cycle or manner where they are re-entrant for the job.
+
+- **Non-deterministic service requirements:** this is where the containers
+and services required to complete the job are not known before the job
+is scheduled and are determined during the execution of the job.
+
+- **Latency:** is the measurement of time between the job being requested
+and initial results returned.
+
+## Workloads Considered
+
+### Trauma / Emergency Case
+
+Immediate diagnostics required for emergency care: immediate
+processing is required for a clinical diagnosis of the trauma. These
+workloads have the patient either awaiting further imaging or
+diagnostics. Seconds count for some emergency diagnostics situations
+(such as stroke).
+
+- Example: Evaluation of Brain Injury in MR/ The typical data size: is
+500 slices / series; 512 x 512 / slice; 16 bits / pixel, Today, the
+typical response time is less than 10 minutes.
+  - This is a case where you have a workload that typically could be done
+asynchronously but now MUST be executed synchronously.
+
+### Scheduled Radiology Workflows
+
+These are often analyzed asynchronously: multiple images/studies are
+analyzed for scheduled clinical studies, research or pro-active
+diagnostics. Typically, in a clinical setting, the imaging has been
+scheduled and the patient will arrive at a predictable time so
+workload can be forecasted.
+
+- Example: Evaluation of Liver in CT; The typical Data Size: 1000 - 2000
+slices in a series; 512 x 512 / slice; 16 bits / pixel. Today, the
+typical response time is greater than 30 minutes.
+- Example: Retrospective analysis over multiple and related data sets
+- Example: existing CT scans could be analyzed when new applications,
+algorithms, or models are available. This would repeat the previous
+example over many patients (100s to 1000s).
+  - This is an example of an asynchronous job that may have a timeout
+measured in tens of minutes.
+
+
+### Quality Verification
+Help the radiology technician improve their results.
+
+- Example: Is it a good mammogram? The typical data size: 2048×2048 /
+image; 12 bits / pixel; the typical response time today is < 2 minutes
+
+- Example: Metadata Quality Assurance : validating the usability of an
+image or data. Does the metadata associated with an image match the
+image? For instance, is this a left or right femur? The label states
+one direction but the image suggests another.
+
+- Example: does the scan need to be repeated? Did the patient move while
+the scan was being made?
+
+### Image Enhancement:
+Using computation and AI techniques, clarify the image for human and machine analysis.
+
+### Medical Text Analysis:
+
+Natural language recognition used with medical notes producing coding
+and diagnostic assistance. For instance, validate that the medical
+regime is both safe and effective for this patient given known medical
+history.
+
+### AI Assisted Labeling (MONAI Label):
+AI assisted labeling is part of the training workflow. MONAI labeling
+assists a human trainer by suggesting areas to be labeled by
+inference. As training progresses, models may be updated for the
+inference engine.
+
+### Speech Based Medical Coding:
+Using natural language recognition, transcribe and code the clinical notes.
+
+## Workloads Out of Scope
+For the purpose of this document, the following workloads are out of scope:
+
+- Genomic Sequence Analysis :
+- Drug Discovery :
+- Training :
+- MONAI Stream : MONAI stream applications. Inference workloads that
+  are being run over live streams of images. An example application
+  would be ultrasound guided medical operations in a clinical
+  setting. This is a highly latency sensitive workload with immediate,
+  synchronous response expected. We expect that the analysis will need
+  to take less than 100ms in order to be useful.
+