In our previous article, Qualtech shared the general product registration requirements for all kinds of artificial intelligence (AI)/machine learning (ML) based software as medical devices (SaMD). Now, TFDA has announced a product registration guidance focusing on AI/ML-based computer-assisted detection (CADe) and computer-assisted diagnosis (CADx) software.
This article summarizes the specific requirements of AI/ML-based CADe and CADx. However, depending on each product's different characteristics, additional documents may still be required during the reviewing process.
|
Items/Technical Documents |
Requirements |
|
Should include: • Possible conditions that may cause device failure or unable to achieve expected performance. • Evaluation of the impact of device-assisted diagnosis (especially the potential risk of false positive and false negative results). • Product usage scenarios, such as simultaneous reading with the user or providing a second read after the user.
• Information, parameters, and data production process of the medical device. The file format of the output result should also be provided. |
|
The algorithm architecture is suggested to include the following items (if applicable): • Name and type of the algorithm. • Development platform, e.g. Tensorflow, Caffe. • Library, e.g. Keras. • Models, e.g. numbers of layers, weights, activation, optimizer, loss function, and metric. • Algorithm functions (including information on image markings and quantitative diagnosis of lesion risk). • Data process flow. |
|
Image processing is suggested to include: • Processing procedures, e.g. filtering, segmentation, normalization, registration, phantom, and artifact or motion correction. • Calibration principles and references of image normalization. |
|
|
Training dataset of the Algorithm should include: • Correspondence of patient population and intended use. During case studies, AI/ML-based SaMD should cover the method in determining disease status, location and affected range, benign/malignant, severity, and stage. • Form, method, and process of data generation. • Additional information, e.g. annotation and clinical diagnosis. |
|
|
Reference standards are suggested to include: • Basis of the reference standard, e.g. predicate device outputs, clinical examination results, biopsy results, or clinicians’ readings. • If the reference standard is based on the reading of clinicians, the number, qualification, experience, and training requirements of the clinicians should be stated. • When multiple clinicians are involved in the formulation of reference standards, the method of combining multiple reading results into one standard should be described. The handling procedure for inconsistent reading results should also be stated. |
|
|
• Detail evaluations of the potential risks and corresponding clinical actions for device failure, unable to achieve expected performance, or false positive/false negative results. • The test dataset must cover as many data source institutions (different regions, different authority institutions) and data process devices (such as different models, types, and production parameters) as possible. • When performing clinical evaluations or clinical trials, statistical analysis should be conducted to compare the independent readings from AI/ML-based CADx medical devices with the traditional readings from clinicians to verify the clinical efficacy of AI/ML-based CADx medical devices in independent diagnosis. |