Data-scarce transfer learning fusion for positron emission tomography and computed tomography guided lung biopsy

OBJECTIVES: We developed a transfer learning-based multimodal fusion deep learning model integrating positron emission tomography/computed tomography (PET/CT) and procedural CT to support CT-guided percutaneous lung lesion biopsy, aiming to overcome PET/CT annotation scarcity and provide automated lesion segmentation and needle trajectory recommendations. METHODS: In this single-center retrospective study of 140 patients, the model combines slice-wise two-dimensional nnU-Net segmentation, dual-stream PET/CT feature extraction, multi-scale cross-attention fusion, source-domain pretraining, and semi-supervised domain adaptation (SSDA). Segmentation was evaluated using Dice similarity coefficient (DSC), 95 % Hausdorff distance, average surface distance, true-positive rate, and positive predictive value. Needle paths were assessed for safety and agreement with manual planning. Ablation studies examined the contribution of model components and different annotation ratios (10 %, 30 %, 50 %). RESULTS: The framework achieved accurate lesion segmentation (mean DSC 0.84 ± 0.06) and maintained high performance with only 30 % annotated slices. Automated needle planning reached 85.7 % success, reduced planning time, and decreased potential complications. Ablation confirmed benefits of pretraining, SSDA, and multimodal fusion. CONCLUSIONS: This framework provides accurate and clinically actionable guidance for CT-guided lung biopsy under annotation-scarce conditions, improving safety, efficiency, and precision in minimally invasive procedures.