PSEUDO-TEXT-CONDITIONED 3D GROUNDING DINO FOR ORGAN LOCALIZATION IN ABDOMINAL CT

Reliable organ localization is a critical prerequisite for automated trauma analysis in abdominal CT. This paper introduces CT-3GDINO, a 3D detector designed to localize the liver, spleen, left kidney, right kidney, and bowel. The authors aim to adapt the query-based Grounding DINO architecture to 3D volumetric data using a semantic-conditioning interface that avoids the need for a full, complex text encoder.

CT-3GDINO utilizes a Swin3D visual backbone to extract hierarchical volumetric features. Instead of using a real text encoder, the model employs frozen, Xavier-initialized pseudo-text class tokens. These tokens are processed by a trainable projection MLP and integrated with visual features through a bidirectional feature enhancer. The model uses language-guided query selection to initialize object queries from semantically relevant image tokens, which are then refined by a cross-modality decoder to predict normalized 3D bounding boxes. The model was trained and evaluated on 193 CT volumes from the RSNA/RATIC dataset, using segmentation-derived boxes as ground truth.

The proposed model achieves an overall top-1 class-wise mAP of 0.5830 across 3D IoU thresholds from 0.1 to 0.7. This performance surpasses both fixed-backbone (0.5570 mAP) and trainable-backbone (0.4657 mAP) classification-pretrained variants. While the model is highly effective at coarse localization (0.9649 AP at IoU 0.1), it shows limited precision for strict box alignment (0.1552 AP at IoU 0.7). The authors conclude that the pseudo-text-conditioned approach provides a robust, open-source baseline for 3D multimodal detection with minimal annotation requirements.