Ischemic heart disease, typified by myocardial infarction and angina pectoris, is a serious group of conditions in which stenosis or occlusion of the coronary arteries impedes blood flow to the myocardium, causing chest pain and myocardial injury. Plaque that forms in the vessel wall as atherosclerosis progresses is the principal cause of this stenosis: as plaque enlarges, the vessel lumen narrows and the blood supply to the myocardium is reduced. Such patients are treated with coronary artery bypass grafting or percutaneous coronary intervention (PCI); of these, PCI---a catheter-based, minimally invasive treatment---has become widely established.
However, PCI carries the risk of complications known as the no-reflow phenomenon and slow flow. No-reflow refers to a state in which, although the stenosis itself is resolved by stent placement or similar procedures, peripheral blood flow is completely obstructed because plaque contents or thrombus released during dilation of the lesion occlude the distal microvasculature (distal microembolization), with microvascular spasm also considered a contributing factor. Slow flow is a state in which blood flow is not completely obstructed but is markedly delayed. These complications are reported to occur in approximately 10--20\% of patients after PCI \citep{babapoor2022}, and some reports indicate an incidence of 10\% or more in cases of acute myocardial infarction (AMI) \textbf{[citation needed]}. Because both cut off the myocardial oxygen supply, they are known to be significant predictors of poor prognosis, including new-onset heart failure, recurrent myocardial infarction, and increased in-hospital mortality \textbf{[citation needed]}. After onset, treatment such as intracoronary administration of vasodilators is performed; however, this cannot remove the lipid and necrotic material that have already occluded the peripheral capillaries, and it also carries the risk of adverse effects such as acute hypotension and arrhythmia, so post-hoc treatment alone has its limits \textbf{[citation needed]}. Against this background, establishing a preventive strategy that forestalls the onset of these complications altogether---rather than addressing them after they occur---has become a critical challenge.
When the risk of no-reflow or slow flow is found to be high before the procedure, clinical practice takes preventive measures such as placing a distal protection device distal to the lesion before stent deployment, or administering vasodilators intracoronarily in advance \textbf{[citation needed]}. While distal protection devices are effective for high-risk lesions, applying them to every case is not realistic given the burden of preparation, making accurate pre-procedural risk stratification essential. Intravascular ultrasound (IVUS) can visualize the condition of the vessel wall and plaque, providing useful information about the lesion. However, interpreting IVUS images requires specialized expertise, and because multiple factors---such as vessel diameter and plaque composition---interact in complex ways, high-precision risk prediction by visual assessment alone is difficult, creating a need for methods that quantitatively assess risk from IVUS images.
Regarding the amount of attenuated plaque observed on IVUS, a significant difference has been reported between patients who developed complications and those who did not \citep{wu2011}. This suggests that a greater amount of attenuated plaque may make thrombus formation and microvascular occlusion more likely, thereby increasing the risk of complications. However, this scoring remains a manual assessment performed by physicians, motivating the need for an automated and objective predictive approach.
Information useful for predicting no-reflow and slow flow is thought to lie at the intersection of the IVUS images themselves, the patient's clinical variables, and the morphology of the lesion. At the same time, because these complications are relatively rare in themselves, and positive cases are even more limited among the cases whose outcomes are labeled, it is difficult to learn all of this information in an end-to-end manner.
Building on the above, this study proposes KSCNet, an architecture that integrally handles IVUS images, patient clinical information, and lesion plaque features. KSCNet is a redesign, adapted to the structure of IVUS data, of a 3D architecture originally developed for kidney stone classification \textbf{[citation needed]}. Because an IVUS pullback is not an isotropic 3D volume like a CT scan but rather an anisotropic sequence of 2D cross-sections arranged along a time series, we adopt, in place of 3D convolution, a weight-shared, per-frame 2D encoder together with sequence modeling via a Transformer across frames. Specifically, it integrates three elements: (1) a mechanism that injects a feature vector---combining patient-level clinical information and plaque features (such as attenuated-plaque area)---as an additional token into the frame sequence, allowing it to interact with the image representation; (2) a mechanism that uses segmentation of the vessel lumen and vessel wall as an auxiliary task and aggregates case-level features via cross-attention, guided by those features; and (3) a multi-task loss that dynamically weights the segmentation loss and the classification loss. This allows us to verify whether combining IVUS images, clinical information, and lesion features contributes to improved predictive performance, compared with methods that use only images, tabular methods that use only clinical information, and existing attenuated-plaque scoring methods.
