Two separate tasks are involved: developing an algorithm for series selection and for the optimal frame selection. For clarification, a single study consists of multiple series, i.e., 2D video files (acquired at different projection angles, and with different settings stored in DICOM metadata) and each series consists of multiple frames. The first task was to extract these relevant for SyntaxScore calculation. This was done by first establishing selection criteria by MD, and then subsequently refined by feedback received from their rejection justifications via dedicated forms incorporated in rich-UI app (Figure 4). A similar approach was used in the case of optimal frame selection. First AI-suggested optimal frame was chosen based on DICOM metadata and image feature analysis. Then the algorithm was iteratively refined by introducing conclusions from MD justifications why the suggested optimal frame was changed, again via dedicated forms incorporated in rich-UI app. The AI-driven workflow significantly reduced the time needed to select the relevant series and optimal frames.

Algorithm development required balancing accuracy with computational efficiency. The first version of an algorithm for coronary artery segmentation was implemented successfully and with very good validation results (DICE score 73%, Accuracy 99%). Next, we provided two significant refinements. The first concerned the introduction of an interobserver variability factor due to the expected complexity of the task. Three experts were asked to perform manual segmentations independently, and as expected their results varied significantly. Embedding this information as a probability map provides additional information to the training process, reducing overfitting and increasing the flexibility of the model. The second refinement concerned the elimination of false positive segmentation results. The first algorithm falsely included no-vessel vessel-like structures i.e. stitches and catheters. We addressed this problem by designing and developing a dedicated classification model. This enabled us to accurately identify vessels and eliminate false positives, resulting in more reliable segmentation results [1]. Through rich-UI interface MD can correct the AI-proposed coronary artery segmentation.

[1] Our paper entitled A novel framework for differentiating vessel-like objects in coronarography images , IEEE EMBC 2023 proceedings.
https://arinex.com.au/EMBC/pdf/full-paper_1180.pdf 

The branch identification algorithm automatically identifies specific vessel branches and names them in accordance with their anatomical descriptions. The algorithm consists of several steps, including the detection of vessel centerlines, the representation of coronary arteries in a graph-like manner, the detection of vessel bifurcations, and finally the identification of vessel branches. The quality of each subsequent step is influenced by the quality of the previous step. For this reason, algorithm steps were developed using an iterative loop refinement approach. As a result, the development efforts led to successful algorithm development and validation. Through a rich-UI interface MD can correct the AI proposed segment identification. 

The reconstruction is performed based on 2D projections acquired during coronarography examination, each in principle at different projection angles and with different settings stored in DICOM metadata. The developed algorithm incorporates geometrical properties of the 2D images, and the information derived from branch identification (i.e. coordinates of the transition between successive segments identified in different projections, centreline coordinates, branch identifiers). First, the 2D images are transformed into the 3D coordinate system, next 3D vessel centerline is constructed by interpolating coordinates derived from single transformed images and taking into consideration the number of projections available per each segment, then the local vessel radius is calculated to maintain a realistic representation of various vessel parts including stenosis, finally a mesh is calculated, and 3D reconstruction is achieved. 

The stenosis identification algorithm makes use of centerline points obtained during step A3, and local radius disappearances. Both geometrical and image features are next compared to the model features to detect any local variances and verify whether a particular centerline point is part of the stenosis or not. Ultimately, the stenosis length can be determined. The algorithm for stenosis characterization is based on a thorough analysis of parameters included in SyntaxScore calculation: qualitative (e.g. thrombosis, tortuosity, calcification), geometrical (e.g. stenosis length, bifurcation angle) and anatomical (e.g. segments involved in stenosis, next visible segment after a stenosis). To categorize stenosis qualitatively, a model based on image features is used. Hand-crafted features were designed and selected to feed the algorithm. To determine stenosis geometric characterization local geometric dependencies derived from previous steps are used. To provide anatomical dependencies, results of segment identification are used. As a result. the algorithm for stenosis identification and characterization was developed. Through rich-UI interface MD can correct the AI proposed stenosis characterization. Their feedback is collected through rich-UI and can be used for further model training.

To develop the Syntax Score calculation algorithm, an extensive analysis of the necessary parameters was performed. This included identifying them and finding a way to generate them from the available algorithms and data. It was necessary to understand some of the parameters more deeply. The algorithm for SyntaxScore calculation takes into account stenosis characterization scored depending on its location determined by branch identification. Each stenosis is located and characterized on at least two independent projections. The proposed algorithm complies with discrepancies in their descriptions by selecting the most severe values. Since SyntaxScore calculation is a complex task that requires medical interpretation and experience in clinical coronarography data analysis, the algorithm development process was conducted in close collaboration with MD. The algorithm was iteratively refined until we covered all differences between AI interpretation and MD scores. As a result, the calculated SyntaxScore value is exactly the same in each instance verified.