A Native Strategy for Integrating Deep-Learning Models for Segmentation into a Radiological Viewer

Abstract

The use of deep-learning (DL) models to support and automate medical imaging diagnostic procedures has become an ongoing focus of research and development. Despite advances in the subject, the integration of such solutions into clinical diagnostic workflows remains challenging. Especially focused on end users, the integration of image-based diagnostic functionalities and access to DL models in a single framework is key to ensuring clinical adoption and usability. This paper proposes a native integration strategy that enables the direct use of DL segmentation models within a CE-marked open-source DICOM viewer without relying on external software, containerised environments, or complex APIs. Unlike previous approaches, which often require technical expertise or infrastructure overhead, the proposed method embeds the model execution pipeline directly into the viewer via a dedicated DL module, maintaining compatibility with clinical standards and allowing model parameters to be set directly from the interface or via a configuration file. To validate the feasibility and versatility of this native integration strategy, two use cases are implemented using models trained in different DL libraries: vertebral bodies segmentation and liver segmentation. The approach proves compatible with heterogeneous model architectures, requires minimal user interaction, and preserves clinical usability without disrupting existing workflows. A new DL integration methodology is presented that combines simplicity, flexibility, and clinical readiness. The proposed framework represents a significant step towards standardised, viewer-native deployment of DL tools, facilitating their adoption in regulated healthcare environments and enabling efficient sharing and reuse of DL models across institutions.

Ähnliche Arbeiten