Creation of Latent Space for Patient Deviation Detection and Similar Nosology Detection Using Unsupervised Metric Learning for Chest Computed Tomography (CT) Images

The developed artificial intelligence models effectively solve specialized medical tasks using methods tailored to specific diseases. However, the diverse anatomical structures inherent in various pathologies pose limitations on the comprehensive assessment of a patient's overall condition using this approach. To overcome this constraint, it is imperative to create an algorithm capable of assessing the general state of the human thorax, thus enabling the detection of deviations that may evade reveal by algorithms focused solely on specific pathologies. Given the considerable data requirements for constructing a highly generalizable model, this study employs a method that enables the use of medical data without annotations. The proposed solution involves training the model using unsupervised metric learning, which will allow for obtaining a hidden space consisting of embeddings of medical images. The aim of this research is to achieve a precise representation of real-world data in the latent space and select a metric that facilitates the identification of patients with similar pathologies. The main challenge in creating such a solution lies in directing the model's attention specifically to nosologies rather than other attributes that may allow for distinguishing between patients, such as primary and secondary sexual characteristics.