This paper presents a novel segmentation method aiming at Hybrid Traumatic Brain Injury Lesion Segmentation. The proposed model was submitted to the MICCAI’s Automated Identification of Moderate-Severe TBI Lesions 2025 (AIMS-TBI25) on behalf of the team "SüSü". The proposed segmentation algorithm comprises a custom V-NET model and filtering based on connected component features. The neural network model was trained using 553 annotated T1-weighted MRI images provided by the challenge organizers. The dataset was split into 455 training and 97 validation images. The learning process was controlled by early stopping to avoid overfitting. Preprocessing normalized intensities to [0, 1] per image and padded smaller scans to \(256 \times 256 \times 256\) voxels. The proposed method couples a 3D VNET-based segmentation network with a lightweight connected component filter that removes small predictions; the voxel number threshold is selected on the validation set via a simple sweep. Training used BCEWithLogitsLoss, Adam optimizer, and Instance Normalization (superior to BatchNorm at batch size 1), with augmentation (flips, \(\pm 30^\circ \) rotations, and random resized crops). Models were trained on the Komondor HPC (A100, \(<\!40\)  GB VRAM) and evaluated under a 16 GB VRAM constraint. In the hidden test set ( \(n{=}223\) ), our submission was ranked 14th on the leaderboard, corresponding to 9th place among unique teams. A deployment oversight caused the algorithm to return empty masks for 31 scans whose size exceeded 256 in at least one dimension; for transparency, we report results on all test scans (mean Dice 0.496) and on the subset where the model actually produced predictions ( \(n{=}192\) , mean Dice 0.571). Despite its simplicity, the connected component filter in the post-processing pipeline yielded an average Dice improvement of 0.098 on the entire test dataset and 0.114 on the filtered dataset compared to raw model predictions.