Introduction Brain metastases from lung cancer typically present as multiple small lesions

creating considerable challenges for accurate segmentation. While existing datasets and models have primarily focused on larger metastases from various primary cancers

there remains a pressing need for tools optimized for small-volume lesions. Therefore

this study aimed to develop and validate a deep learning model specifically designed for segmenting small-volume brain metastases originating from lung cancer. Materials and methods We collected 1413 magnetic resonance imaging (MRI) scans (from two institutions) containing brain metastases with a median volume of 32 mm3 (interquartile range 103 mm3)

substantially smaller than the median volume found in the BraTS-METS dataset. The dataset included gradient-echo T1-weighted contrast-enhanced (61.9%) and black-blood sequences (38.1%). We modified nnU-Net by incorporating focal loss ( λ = 0.5) in addition to standard Dice and cross-entropy losses to improve small lesion detection. Post-processing using FreeSurfer skull stripping was applied to reduce false positives in skull regions. The model was evaluated on internal (n = 283) and external (n = 373) test sets. Results The proposed model achieved a mean Dice similarity coefficient of 0.7416 ± 0.2694 on the internal test set and 0.7587 ± 0.2878 on the external dataset for lesions >100 mm3. Ablation studies indicated that focal loss improved performance over baseline nnU-Net

while skull stripping further reduced false positives. Conclusions The developed model demonstrates reliable segmentation of small-volume brain metastases from lung cancer across different MRI sequences and scanner manufacturers. This approach may reduce treatment planning time and improve consistency in target delineation for stereotactic radiosurgery.