What question did this study set out to answer?

The aim is to develop a model for visual grounding of findings in PET/CT images using an automated weak-labeling pipeline.

ConTEXTual Net 3D: Vision-Language Modeling in PET/CT for Visual Grounding of Positive Findings

View Full Paper

Ask AI

Bookmark

The aim is to develop a model for visual grounding of findings in PET/CT images using an automated weak-labeling pipeline.
Developed an automated pipeline for generating weak image-text labels from PET/CT reports.
Identified sentences with positive findings using standardized uptake values and axial slice numbers.
Extracted 11,356 sentence-label pairs from 25,578 PET/CT exams for model training.
Trained ConTEXTual Net 3D on the generated dataset and evaluated its performance with F1 scores.
Weak labeling pipeline accurately identified lesion locations in 98% of cases.
ConTEXTual Net 3D achieved an F1 score of 0.80, outperforming LLMSeg (0.22) and the 2.5D model (0.53).
Both radiologists had higher F1 scores (0.94 and 0.91) compared to ConTEXTual Net 3D.
Better performance observed on certain exams like <sup>18</sup>F-fluorodeoxyglucose (F1 = 0.78) than on DOTATATE (F1 = 0.58).

No takes yet. Share an insight, caveat, or question.

Automated pipeline generates annotated image-text pairs in PET/CT, suggesting improved radiology reporting capabilities.

Huemann et al. (2026) studied this question.

View Full Paper

Ask AI

Bookmark