State-of-the-art models for medical image segmentation achieve excellent accuracy but require substantial computational resources, limiting deployment in resource-constrained clinical settings. We present SegMate, an efficient 2.5D framework that achieves state-of-the-art accuracy, while considerably reducing computational requirements. Our efficient design is the result of meticulously integrating asymmetric architectures, attention mechanisms, multi-scale feature fusion, slice-based positional conditioning, and multi-task optimization. We demonstrate the efficiency-accuracy trade-off of our framework across three modern backbones (EfficientNetV2-M, MambaOut-Tiny, FastViT-T12). We perform experiments on three datasets: TotalSegmentator, SegTHOR and AMOS22. Compared with the vanilla models, SegMate reduces computation (GFLOPs) by up to 2.5x and memory footprint (VRAM) by up to 2.1x, while generally registering performance gains of around 1%. On TotalSegmentator, we achieve a Dice score of 93.51% with only 295MB peak GPU memory. Zero-shot cross-dataset evaluations on SegTHOR and AMOS22 demonstrate strong generalization, with Dice scores of up to 86.85% and 89.35%, respectively. We release our open-source code at https://github.com/andreibunea99/SegMate.
Figures & tables
Figure 1: Segmentation accuracy vs. GFLOPs per volume ( ≈ 162 slices) on TotalSegmentator. Point size indicates VRAM usage. All SegMate variants are below 300 MB.
Figure 2: SegMate architecture. CBAM modules are positioned inside the main decoder blocks after encoder-decoder fusion; SE gates nested skip blocks that cascade cross-scale features. FiLM conditions the bottleneck on normalized slice position znorm .
Backbone
Method
TotalSeg
ST ZS
ST FT
A22 ZS
A22 FT
VRAM
GFLOPs
EffNetV2-M
SegMate
93.51
86.85
88.51
89.35
91.29
295 MB
5,097
Vanilla
92.45
85.36
88.06
89.29
91.22
374 MB
6,034
MambaOut-T
SegMate
92.42
86.46
86.67
89.14
90.44
208 MB
4,384
Vanilla
91.16
83.11
86.66
88.22
90.28
431 MB
9,905
FastViT-T12
SegMate
92.25
84.23
87.13
88.87
91.52
120 MB
1,476
Vanilla
92.75
83.19
86.02
88.24
88.51
253 MB
3,688
Table 1: Dice scores (%) over all organs, across three datasets. For SegTHOR (ST) and AMOS22 (A22), we show zero-shot (ZS) and fine-tuning (FT) transfer from TotalSegmentator. Per-volume GFLOPs for a 512×512×162 CT volume with 162 sequential slices. Best per-backbone is in bold . Scores represent averages over 3 runs.
Method
Type
ZS Dice
FT Dice
Train Set
TotalSegmentator
SegMate (EffNetV2-M)
2.5D
–
93.51
TotalSeg
nnU-Net [ 17 ]
2D
–
94.99
TotalSeg
Standard U-Net
2D
–
87.48
TotalSeg
SegTHOR
SegMate (EffNetV2-M)
2.5D
86.85
88.51
TotalSeg (ZS) / SegTHOR (FT)
Table 2: Comparison with state-of-the-art methods across three datasets (3D volume-level Dice %). ZS: zero-shot transfer from TotalSegmentator; FT: fine-tuned. † Models evaluated by PaR [ 15 ] . ‡ Models evaluated by Bio2Vol [ 40 ] . Best per dataset is in bold .
Measure
Vanilla
+Asym. dec.
+CBAM
+SE
+2.5D
+Slice Pos.
Dice (%) ↑
92.45
+0.45
+0.65
+0.81
+0.95
+1.06
HD95 (mm) ↓
4.97
− 0.62
− 0.61
− 0.80
− 0.81
− 0.84
VRAM (MB) ↓
373.9
− 85.5
− 85.5
− 85.4
− 85.4
− 85.5
GFLOPs/slice ↓
37.25
− 5.79
− 5.78
− 5.78
− 5.78
− 5.78
Table 3: SegMeta ablation for EfficientNetV2-M applied on TotalSegmentator. All results are reported at 2D slice level (organ-averaged). Values after vanilla baseline show absolute percentage point change w.r.t the baseline. Highest gain is in bold .