Development and validation of a deep learning-based fully automated algorithm for pre-TAVR CT assessment of the aortic valvular complex and detection of anatomical risk factors: a retrospective, multicentre study

Figures

• 

  Fig. 1. Algorithm development and validation flowchart. (a) Represents the algorithm development schematic, which includes DL model training and core function algorithm workflow (Number of training dataset = 800). (b) Represents the algorithm validation schematic, which includes algorithm performance verification, algorithm function verification, and algorithm efficiency verification (Number of internal/external validation datasets = 452/100).

• 

  Fig. 2. FORSSMANN assessment result and 3D reconstruction. (a) By taking the TAV as an example, the anatomical structure of the AVC region and the basic planes that need to be located and measured. The aortic angle was determined from the annular plane. The AA plane was located at the widest part of the AA. The STJ plane was located at the narrowest position where the aortic sinus transitions to the AA. The SOV plane was identified as the plane with the largest cross-sectional area of the lumen, based on the centreline between the annulus and STJ. In addition, the plane 4 mm below the annular plane vertically is the left ventricular outflow tract (LVOT) plane required for pre-operative evaluation of TAVR. The accuracy of LVOT positioning directly depends on the annular plane; therefore, this study did not conduct separate verification for the LVOT plane. (b) The algorithm completes the 3D reconstruction of the AVC model after automatic segmentation, localisation, and measurement. In the 3D reconstructed model, the aorta is shown in purple, the right coronary artery in red, the left coronary artery in green, the non-coronary cusp in orange, the right coronary cusp in blue, the left…

• 

  Fig. 3. Schematic illustration of methods for positioning the annular plane in different anatomical classifications of the aortic valve. (a) For tri-leaflet structures (TAV, Type 1 BAV, Type 2 BAV), the annular plane is located at the plane determined by all three of the most basal attachment points of the aortic valve cusps. (b) For bi-leaflet structures (Type 0 BAV), the annular plane was positioned tangent to both the basal hinge points to achieve the smallest cross-sectional lumen.

• 

  Fig. 4. Statistical comparison of Euclidean distance for spatial localisation of specific anatomical structures by FORSSMANN and senior observers. (a) Statistical analysis results for the internal validation dataset (n = 452), (b) Statistical results for the external validation dataset (n = 100).

• 

  Fig. 5. BA plotting comparing senior observers and the FORSSMANN algorithm. Comparison of (a) APD, (b) AA, (c) SOV, (d) STJ, (e) COH Left, (f) COH Right, (g) VOC, and (h) aoric angle with the y-axis representing the difference between senior observers and the FORSSMANN algorithm while the x-axis representing the average between these two values. The area covered by the blue shadow represents the confidence interval of the internal validation dataset (n = 452), while the area covered by the red shadow represents the confidence interval of the external validation dataset (n = 100).

• 

  Fig. 6. The impact of severe calcification and different aortic valve anatomy on FORSSMANN. (a) The impact of severe calcification on FORSSMANN in the internal and external validation datasets. (b) The impact of different valve leaflet types on FORSSMANN in the internal and external validation datasets (internal validation dataset n = 452) (external validation dataset n = 100).

• 

  Fig. 7. The measurement differences of the APD. Comparing the performance of the FORSSMANN algorithm and junior observers, with the senior observers serving as the reference standard (internal validation dataset n = 452) (external validation dataset n = 100).

• 

  Fig. 8. Risk factor detection results. The plotting figures of FORSSMANN evaluation value verse senior observers of (a) horizontal aorta, (b) AA dilation, (c) severe calcification, (d) exceptional SOV sizing, (e) low left COH and (f) low right COH (TP = True Positive, TN = True Negative, FP = False Positive, FN = False Negative, NA = Not Applicable). “Positive” refers to the detection of anatomical risk by senior observers, whereas “Negative” indicates no detection. “True” represents a match in detection results between FORSSMANN and senior observers, while “False” indicates a mismatch in detections (n = 552).