Related Experiment Video
Updated: Jul 12, 2026
04:25
Manual Segmentation of the Human Choroid Plexus Using Brain MRI
Published on: December 15, 2023
Integrating 3D Volumetric Segmentation and LLM-Based Classification for csPCa Detection on mpMRI: Multi-Institutional
Kexin Wang1, Ge Gao1, Huihui Wang1
1Department of Radiology, Peking University First Hospital, Beijing, China (K.W., G.G., H.W., J.W., X.W.).
Academic Radiology
|July 9, 2026
Summary
Combining 3D segmentation and large language models (LLMs) significantly improves prostate cancer detection on MRI. These integrated models offer enhanced accuracy and clinical utility over single approaches for clinically significant prostate cancer (csPCa).
Area of Science:
- Artificial Intelligence in Medical Imaging
- Oncology
- Radiology
Background:
- Clinically significant prostate cancer (csPCa) detection on multiparametric Magnetic Resonance Imaging (mpMRI) remains challenging.
- Existing methods often rely on single-modality analysis, potentially limiting diagnostic accuracy.
- Integration of advanced AI techniques may enhance csPCa detection.
Purpose of the Study:
- To develop and validate integrated models combining 3D volumetric segmentation and LLM-based slice-wise classification for csPCa detection.
- To compare the performance of integrated models against individual segmentation and LLM approaches.
- To assess the clinical utility and generalizability of the combined models in an external validation cohort.
Main Methods:
- Retrospective multi-institutional study with 5050 patients (3896 development, 1154 external validation).
- A 3D V-Net model for voxel-wise csPCa segmentation and a MedGemma-IT LLM for slice-level classification were developed.
- Two combined models integrated V-Net and LLM metrics; performance evaluated using AUROC, AUPRC, DCA, NRI, and IDI.
Main Results:
- Combined Model 1 (AUROC 0.900) and Combined Model 2 (AUROC 0.885) significantly outperformed individual V-Net and LLM metrics in external validation.
- Both combined models showed significant improvements in NRI and IDI compared to single-modality approaches.
- Decision curve analysis indicated superior standardized net benefit for combined models across clinically relevant risk thresholds.
Conclusions:
- Integration of 3D volumetric segmentation and LLM-based classification substantially enhances csPCa detection accuracy and clinical utility.
- The developed combined models demonstrate promising generalizability across diverse mpMRI scanners and protocols.
- This hybrid AI approach represents a significant advancement over single-modality methods for prostate cancer diagnosis.
