Jianlin Jack Cheng

Cell membrane segmentation in cryo-electron tomograms now runs through a two-stage fine-tuned pipeline: tomograms are sliced along Z and treated as video frames, and SAM2's memory encoder and memory attention track membranes across slices. Stage one produces an initial membrane prediction whose normalized logits are concatenated with the original slices before stage two emits the final mask. On 17 independent experimental tomograms with ground truth, Dice is 0.8116, IoU 0.7059, precision 0.8123 and recall 0.8430, all above Membrain-Seg and TARDIS, with AUPRC 0.8565; average runtime is 79.97 seconds per tomogram at 1312.1 MB of GPU memory.

Cell membrane segmentation in cryo-electron tomograms now runs through a two-stage fine-tuned pipeline: tomograms are sliced along Z and treated as video frames, and SAM2's memory encoder and memory attention track membranes across slices. Stage one produces an initial membrane prediction whose normalized logits are concatenated with the original slices before stage two emits the final mask. On 17 independent experimental tomograms with ground truth, Dice is 0.8116, IoU 0.7059, precision 0.8123 and recall 0.8430, all above Membrain-Seg and TARDIS, with AUPRC 0.8565; average runtime is 79.97 seconds per tomogram at 1312.1 MB of GPU memory.