Multiparametric MRI deep learning model differentiates lung from non-lung brain metastases

A multiparametric MRI-based 3D deep learning fusion model differentiated lung cancer from non-lung cancer brain metastases with an AUC of 0.925, outperforming single-sequence models (0.817, 0.802). Patient-level AUC was 0.942. Radiologist accuracy improved with AI assistance (tr…

• The 3D fusion model using CE-T1 and T2-FLAIR sequences with habitat-based subregion analysis and attention mechanisms achieved lesion-level AUC 0.925.
• Single-modality models (CE-T1 AUC=0.817; T2-FLAIR AUC=0.802) were significantly outperformed; the model was robust across histologic subtypes (AUCs 0.895-0.956).
• A multi-reader study showed AI assistance improved accuracy for all radiologist levels: trainees 0.500→0.583, experienced 0.595→0.679, experts 0.690→0.845.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.