.jpg)
Segmentation of vascular structures in computed tomography is a critical task for diagnosis and surgical planning. However, manual annotations often present significant noise, especially when each volume is labeled only once. This problem, known as single-mask noise, affects the quality of machine learning models and hinders error auditing. Recent research proposes an approach based on cross-sectional patch self-consistency to detect conflicting regions without needing to merge multiple annotators. The idea is that, along blood vessels, patches extracted perpendicularly to the centerline exhibit anatomical recurrence: similar patches in intensity should have coherent masks. When statistical discrepancies are found in the labels of those equivalent patches, a potentially noisy annotation is identified.
This method, being decoupled from network training, allows explicit auditing and generates quality maps per scan. Applied to coronary angiographies, it reveals systematic biases: transverse and oblique vessels show error rates up to five times higher than those aligned with the axis, in addition to correlations with cross-sectional area and intensity. Detecting and correcting these biases is essential to improve the robustness of clinical models.
The practical implementation of these workflows requires a solid technological infrastructure. At Q2BSTUDIO, we develop AI for businesses that integrate computer vision and patch processing techniques. Our experience in custom applications allows us to build modular pipelines that address everything from DICOM data ingestion to quality report generation. Additionally, we combine artificial intelligence with AWS and Azure cloud services to scale the analysis of large volumes of images without compromising security, a critical aspect in healthcare environments where cybersecurity is a priority.
Our team also offers business intelligence services using Power BI to visualize annotation quality metrics and biases, facilitating decision-making. Likewise, we develop AI agents that automate noise detection and suggest corrections, reducing the workload of radiologists. All of this materializes in custom software that adapts to the specific needs of each institution.
Noise detection in annotations not only improves model performance but also allows auditing the labeling process and correcting systematic biases. With tools like those we offer at Q2BSTUDIO, diagnostic centers can elevate the quality of their data and, ultimately, the accuracy of their diagnoses assisted by artificial intelligence.
