(153 days)
Saige-Dx analyzes digital breast tomosynthesis (DBT) mammograms to identify the presence or absence of soft tissue lesions and calcifications that may be indicative of cancer. For a given DBT mammogram, Saige-Dx analyzes the DBT image stacks and the accompanying 2D images, including full field digital mammography and/or synthetic images. The system assigns a Suspicion Level, indicating the strength of suspicion that cancer may be present, for each detected finding and for the entire case. The outputs of Saige-Dx are intended to be used as a concurrent reading aid for interpreting physicians on screening mammograms with compatible DBT hardware.
Saige-Dx is a software device that processes screening mammograms using artificial intelligence to aid interpreting radiologists. By automatically detecting the presence or absence of soft tissue lesions and calcifications in mammography images, Saige-Dx can help improve reader performance, while also reducing time. The software takes as input a set of x-ray mammogram DICOM files from a single digital breast tomosynthesis (DBT) study and generates finding-level outputs for each image analyzed, as well as an aggregate case-level assessment. Saige-Dx processes both the DBT image stacks and the associated 2D images (full-field digital mammography (FFDM) and/or synthetic 2D images) in a DBT study. For each image, Saige-Dx outputs bounding boxes circumscribing any detected findings and assigns a Finding Suspicion Level to each finding, indicating the degree of suspicion that the finding is malignant. Saige-Dx uses the results of the finding-level analysis to generate a Case Suspicion Level, indicating the degree of suspicion for malignancy across the case. Saige-Dx encapsulates the finding and case-level results into a DICOM Structured Report (SR) object containing markings that can be overlaid on the original mammogram images using a viewing workstation and a DICOM Secondary Capture (SC) object containing a summary report of the Saige-Dx results.
Here's a breakdown of the acceptance criteria and the study proving the device meets them, based on the provided text:
1. A table of acceptance criteria and the reported device performance
| Acceptance Criteria (Endpoint) | Reported Device Performance |
|---|---|
| Substantial equivalence demonstrating non-inferiority of the subject device (Saige-Dx) on compatible exams compared to the predicate device's performance on previously compatible exams. | The study endpoint was met. The lower bound of the 95% CI around the delta AUC between Hologic and GE cases, compared to Hologic-only exams, was greater than the non-inferiority margin. |
| Case-level AUC on compatible exams: 0.910 (95% CI: 0.886, 0.933) | |
| Generalizable standalone performance across confounders for GE and Hologic exams. | Demonstrated generalizable standalone performance on GE and Hologic exams across patient age, breast density, breast size, race, ethnicity, exam type, pathology classification, lesion size, and modality. |
| Performance on Hologic HD images. | Met pre-specified performance criteria. |
| Performance on unilateral breasts. | Met pre-specified performance criteria. |
| Performance on breast implants (implant displaced views). | Met pre-specified performance criteria. |
2. Sample size used for the test set and the data provenance
- Sample Size: 1,804 women (236 cancer exams and 1,568 non-cancer exams).
- Data Provenance: Collected from 12 clinical sites across the United States. It's a retrospective dataset, as indicated by the description of cancer exams being confirmed by biopsy pathology and non-cancer exams by negatively interpreted subsequent screens.
3. Number of experts used to establish the ground truth for the test set and the qualifications of those experts
- Number of Experts: At least two independent truthers, plus an additional adjudicator if needed (implying a minimum of two, potentially three).
- Qualifications of Experts: MQSA qualified, breast imaging specialists.
4. Adjudication method for the test set
- Adjudication Method: "Briefly, each cancer exam and supporting medical reports were reviewed by two independent truthers, plus an additional adjudicator if needed." This describes a 2+1 adjudication method.
5. If a multi reader multi case (MRMC) comparative effectiveness study was done, If so, what was the effect size of how much human readers improve with AI vs without AI assistance
- The provided text describes a standalone performance study ("The pivotal study compared the standalone performance between the subject device"). It does not mention an MRMC comparative effectiveness study and therefore no effect size for human reader improvement with AI assistance is reported. The device is intended as a concurrent reading aid, but the reported study focused on the algorithm's standalone performance.
6. If a standalone (i.e. algorithm only without human-in-the-loop performance) was done
- Yes, a standalone performance study was done. The text states: "Validation of the software was performed using standalone performance testing..." and "The pivotal study compared the standalone performance between the subject device."
7. The type of ground truth used
- For Cancer Exams: Confirmed by biopsy pathology.
- For Non-Cancer Exams: Confirmed by a negatively interpreted exam on the subsequent screen and without malignant biopsy pathology.
- For Lesions: Lesions for cancer exams were established by MQSA qualified breast imaging specialists, likely based on radiological findings and pathology reports.
8. The sample size for the training set
- Sample Size: 121,348 patients and 122,252 studies.
9. How the ground truth for the training set was established
- The document does not explicitly detail the method for establishing ground truth for the training set. It mentions the training dataset was "robust and diverse." However, given the rigorous approach described for the test set's ground truth (biopsy pathology, negative subsequent screens, expert review), it is reasonable to infer a similar, if not identical, standard was applied to the training data. The text emphasizes "no exam overlap between the training and testing datasets," indicating a careful approach to data separation.
§ 892.2090 Radiological computer-assisted detection and diagnosis software.
(a)
Identification. A radiological computer-assisted detection and diagnostic software is an image processing device intended to aid in the detection, localization, and characterization of fracture, lesions, or other disease-specific findings on acquired medical images (e.g., radiography, magnetic resonance, computed tomography). The device detects, identifies, and characterizes findings based on features or information extracted from images, and provides information about the presence, location, and characteristics of the findings to the user. The analysis is intended to inform the primary diagnostic and patient management decisions that are made by the clinical user. The device is not intended as a replacement for a complete clinician's review or their clinical judgment that takes into account other relevant information from the image or patient history.(b)
Classification. Class II (special controls). The special controls for this device are:(1) Design verification and validation must include:
(i) A detailed description of the image analysis algorithm, including a description of the algorithm inputs and outputs, each major component or block, how the algorithm and output affects or relates to clinical practice or patient care, and any algorithm limitations.
(ii) A detailed description of pre-specified performance testing protocols and dataset(s) used to assess whether the device will provide improved assisted-read detection and diagnostic performance as intended in the indicated user population(s), and to characterize the standalone device performance for labeling. Performance testing includes standalone test(s), side-by-side comparison(s), and/or a reader study, as applicable.
(iii) Results from standalone performance testing used to characterize the independent performance of the device separate from aided user performance. The performance assessment must be based on appropriate diagnostic accuracy measures (
e.g., receiver operator characteristic plot, sensitivity, specificity, positive and negative predictive values, and diagnostic likelihood ratio). Devices with localization output must include localization accuracy testing as a component of standalone testing. The test dataset must be representative of the typical patient population with enrichment made only to ensure that the test dataset contains a sufficient number of cases from important cohorts (e.g., subsets defined by clinically relevant confounders, effect modifiers, concomitant disease, and subsets defined by image acquisition characteristics) such that the performance estimates and confidence intervals of the device for these individual subsets can be characterized for the intended use population and imaging equipment.(iv) Results from performance testing that demonstrate that the device provides improved assisted-read detection and/or diagnostic performance as intended in the indicated user population(s) when used in accordance with the instructions for use. The reader population must be comprised of the intended user population in terms of clinical training, certification, and years of experience. The performance assessment must be based on appropriate diagnostic accuracy measures (
e.g., receiver operator characteristic plot, sensitivity, specificity, positive and negative predictive values, and diagnostic likelihood ratio). Test datasets must meet the requirements described in paragraph (b)(1)(iii) of this section.(v) Appropriate software documentation, including device hazard analysis, software requirements specification document, software design specification document, traceability analysis, system level test protocol, pass/fail criteria, testing results, and cybersecurity measures.
(2) Labeling must include the following:
(i) A detailed description of the patient population for which the device is indicated for use.
(ii) A detailed description of the device instructions for use, including the intended reading protocol and how the user should interpret the device output.
(iii) A detailed description of the intended user, and any user training materials or programs that address appropriate reading protocols for the device, to ensure that the end user is fully aware of how to interpret and apply the device output.
(iv) A detailed description of the device inputs and outputs.
(v) A detailed description of compatible imaging hardware and imaging protocols.
(vi) Warnings, precautions, and limitations must include situations in which the device may fail or may not operate at its expected performance level (
e.g., poor image quality or for certain subpopulations), as applicable.(vii) A detailed summary of the performance testing, including test methods, dataset characteristics, results, and a summary of sub-analyses on case distributions stratified by relevant confounders, such as anatomical characteristics, patient demographics and medical history, user experience, and imaging equipment.