arXiv:2606.24570v1 Announce Type: new 
Abstract: Vision-language contrastive pretraining has become the dominant recipe for 3D medical foundation models, leveraging the large volumes of paired scans and reports produced in clinical practice. However, medical images usually span dozens of organs, and radiological reports are much longer than typical natural image captions and are composed of multiple structured sections. CLIP-style pretraining compresses this structure by encoding each modality into a single global token, at the risk of losing important details. We introduce ConQuer (Concept Queries), an image-text pretraining method that augments CLIP's global alignment with a set of localized alignments, one per concept. ConQuer splits the report into concept-specific sections and learns cross-attention queries that pool the matching image features without using any segmentation mask or spatial supervision. Contrastive learning is then applied independently for each concept. Concepts can be any unit of semantic localization; here, they are anatomical regions, one query per organ or gross body region. As a byproduct, each query learns attention maps focused on its concept, providing built-in spatial interpretability. We use ConQuer to train Jolia, a 3D CT foundation model on chest and abdominal CT. Jolia consistently outperforms a CLIP baseline on findings classification, report generation, and cross-center transfer, and sets a new state of the art across multiple public benchmarks. Jolia's weights will be released upon acceptance.

تم تقديم إطار عمل جديد يسمى ConQuer لتحسين التدريب المتباين بين الرؤية واللغة لنماذج الأساس الطبية ثلاثية الأبعاد، حيث يعالج القيود الموجودة في الأساليب الحالية من خلال السماح بالتوافقات المحلية لكل مفهوم في التقارير الإشعاعية. تهدف هذه الطريقة إلى الحفاظ على التفاصيل المهمة التي غالبًا ما تضيع في التدريب التقليدي على نمط CLIP.

Se ha introducido un nuevo marco llamado ConQuer para mejorar el preentrenamiento contrastivo de visión-lenguaje para modelos de fundación médica en 3D, abordando las limitaciones de los métodos existentes al permitir alineaciones localizadas para cada concepto en los informes radiológicos. Este enfoque busca preservar detalles cruciales que a menudo se pierden en el preentrenamiento tradicional de estilo CLIP.

Un nouveau cadre nommé ConQuer a été introduit pour améliorer le préentraînement contrastif vision-langage pour les modèles de fondation médicale en 3D, abordant les limitations des méthodes existantes en permettant des alignements localisés pour chaque concept dans les rapports radiologiques. Cette approche vise à préserver des détails cruciaux souvent perdus dans le préentraînement traditionnel de type CLIP.

A new framework named ConQuer has been introduced to enhance vision-language contrastive pretraining for 3D medical foundation models, addressing the limitations of existing methods by allowing localized alignments for each concept in radiological reports. This approach aims to preserve crucial details that are often lost in traditional CLIP-style pretraining.