Table 1 Checklist of steps required for management of project involving deep learning
Scope | â Define scope of project: detection, segmentation, classification, monitoring, prediction or prognosis. |
|---|---|
Team building | â Project manager (e.g, physician, data scientist) â Clinical expertise (e.g., surgeon or hepatologist) â Imaging expertise (e.g., radiologist) â Technical expertise (e.g., data scientist) |
Ethics | â Obtain IRB approval |
Cohorting | â Selection process (e.g., by target population vs. database) â Definition of eligibility criteria â Identification of data source |
Data | De-identification â Data anonymization vs. pseudonymization Collection and curation â Data collection â Data exploration and quality control â Labeling = markup and annotations â Reference standard (synonyms: ground truth or gold standard) Sampling â Creation of training, validation and test datasets â Alternative: cross-validation |
Model | â Defining performance metrics â Selection of model (convolutional, recurrent, fully connected) and librairies â Running the experiment followed by hyperparameters fine tuning â Testing: assessing performance on separate test dataset |
Hardware | â Determine best configuration based on model architecture and memory requirements â Local (CPUs vs. GPUs) vs. cloud computing (GPUs vs. TPUs) |
Regulatory | â Market research to inform decision to commercialize â Quality management system â Compliance with local regulatory jurisdictions |
Clinical adoption | â Integration in distribution platform â Clinical validation of performance â Deployment in clinical practice |