Parameter-Efficient 3D Segmentation of Liver and Liver tumors: Depthwise factorization Scales Better Than Dense Convolution with Spatial Dimensionality
Authors: Adham M. Alkhadrawi, Mohammed A. B. Mahmoud
Organizations: Department of Molecular Biosciences and Bioengineering, University of Hawaii at Manoa, Honolulu, HI, USA · Faculty of Computer Studies, Arab Open University, Riyadh 11681, Kingdom of Saudi Arabia
Three-dimensional dense convolutional networks are the strongest performers on volumetric medical image segmentation, but their parameter counts scale poorly: moving a dense k x k convolution to k x k x k multiplies its weights by k. We observe that depthwise separable factorization does not share this penalty. Because the cubic kernel term applies only to the depthwise stage while the pointwise projection, which dominates the parameter count, is unchanged, the same architecture grows by 5 % from 2D to 3D where a dense convolutional U-Net grows by 200 %. We exploit this asymmetry to build a 3D U-Net with 536,990 parameters, 24x fewer than an identical dense 3D U-Net. On MSD Task03 Liver (the Medical Segmentation Decathlon liver task, derived from LiTS), evaluated per case under five-fold cross-validation over all 131 public volumes, the model reaches a tumor Dice of 0.577 (95% CI [0.518, 0.633]) and a liver Dice of 0.947 (95% CI [0.941, 0.952]). Its liver Dice exceeds previously reported performance. Its tumor Dice exceeds their low-resolution configuration (0.4701) by 0.107 and their 2D configuration (0.5394), at approximately one twenty-fourth of the parameters and roughly half the in-plane resolution. Trained under identical conditions on a common held-out split, it exceeds a dense 3D U-Net on liver by +0.031 Dice (paired p = 0.006) and on tumor by +0.041 (95 % CI [+0.005, +0.081], paired p = 0.056), suggesting the factorization also acts as a regularizer in the small-data regime characteristic of medical imaging. We further show, on both LiTS and a 2D endoscopy benchmark, that a large fraction of the network's learnable spatial filters can be replaced by fixed shifts at no cost in accuracy, but that replacing all of them is measurably worse, the placement of spatial capacity matters more than its total amount.
Figures & tables
Convolution
2D
3D
Growth
Dense kd
17,266,371
51,772,227
3.00×
Depthwise separable
1,978,252
2,073,886
1.05×
Table 1: Parameter count of an identical U-Net topology under 2D and 3D convolution. The depthwise separable variant pays the cubic kernel term only on the depthwise stage; the pointwise projection, which dominates, is unchanged.
Figure 1: The scaling asymmetry, computed from Equation 1 over the four encoder stages of the architecture in Figure 2 . (a) Moving a dense U-Net from k×k to k×k×k kernels multiplies its convolution weights by k . The depthwise separable form does not pay this, because the cubic term applies only to the depthwise stage while the pointwise projection, which dominates the count, is unchanged. (b) The same data as growth rates: 200% against 5% .
Figure 2: (A) The 3D U-Net backbone. Numbers give the channel width at each level; the base width of 32 is halved relative to the 64 conventional in 2D because activation memory, not parameter count, binds in 3D. Dashed arrows are skip connections. (B) The three convolutional block variants compared in this work. All share the structure spatial mixing followed by channel mixing and differ only in the spatial stage, so any difference in accuracy is attributable to that stage alone.
Model
Parameters
Liver Dice
tumor Dice
Ours (hybrid)
536,990
0.947 [0.941, 0.952]
0.577 [0.518, 0.633]
Table 2: MSD Liver per-case Dice under five-fold cross-validation over all 131 public volumes. Every volume is held out exactly once. No checkpoint selection is performed: each fold trains for a fixed 200 epochs and the final weights are scored, because the held-out volumes serve as both the validation and the reported set. Post-processing applied as described in Section 3.6. Intervals are 95% percentile bootstrap over cases ( 20,000 resamples), which is the uncertainty relevant to generalization; see Section 4.2 . Parameter counts are exact.
Method
Params
Liver
tumor
nnU-Net, 2D [ 13 ]
—
0.9437
0.5394
nnU-Net, 3D low-res only [ 13 ]
—
0.9469
0.4701
nnU-Net, 3D cascade [ 13 ]
—
0.9538
0.5849
nnU-Net, 3D [ 13 ]
—
0.9411
0.6174
nnU-Net, 3D + 3D cascade ensemble [ 13 ]
—
0.9543
0.6182
Dense 3D U-Net (this work, single split)
12,946,851
0.9140
0.4578
Table 3: Comparison on MSD Task03_Liver. All rows are five-fold cross-validation over the same 131 public volumes, with no separate held-out set, so the figures are directly comparable. We omit nnU-Net’s challenge-submission entry, which was scored on withheld cases we have no access to. nnU-Net does not report parameter counts for these configurations.
Fold
Volumes
tumor-free
tumor (all)
tumor (bearing)
0
27
3
0.5801
0.5276
1
26
1
0.6543
0.6405
2
26
3
0.5753
0.5634
3
26
6
0.4872
0.4834
4
26
0
0.5892
0.5892
Pooled
131
13
0.577
0.565
Table 4: Per-fold tumor Dice (all-volume LiTS convention). Fold means span 0.487 to 0.654 . A permutation test that reassigns the same 131 per-case scores to folds at random reproduces this spread exactly: the observed standard deviation of 0.0596 sits on the null median of 0.0597 ( p=0.50 ), and a fold as low as 0.487 occurs in 31% of random assignments. The variation between folds is therefore attributable to which cases fell together, not to differences between the trained models.
Figure 3: (a) Per-fold tumor Dice, with the pooled mean (dashed) and its 95% bootstrap interval over cases (shaded). (b) Permutation null: the same 131 per-case scores reassigned to folds of the same sizes at random, 20,000 times. The observed standard deviation across folds falls at the median of that null, so the fold-to-fold spread carries no information about the trained models, it reflects only which cases fell together.
Variant
Parameters
Liver
tumor (all 19)
tumor (17 bearing)
Dense 3D U-Net
12,946,851
0.914
0.458
0.512
(a) Fixed shift
485,987
0.941
0.574
0.524
(b) Learnable depthwise
557,726
0.945
0.596
0.548
(c) Hybrid
536,990
0.945
0.600
0.553
Table 5: Block variant ablation on MSD Liver: one seed per variant, identical training, identical post-processing ( T=50 voxels), evaluated on the same held-out split of 19 volumes. Comparisons against the dense baseline are paired over volumes. A single seed is used for every row so that the paired structure is preserved; the three-seed figures for the hybrid variant are reported separately in Section 4.9 .
Seed
Encoder
Deep decoder
Shallow decoder
0
5.47
1.94
3.80
1
5.66
2.04
3.87
2
5.47
2.17
3.90
Table 6: Mean number of open taps per depthwise filter under L0 gating ( λ=10−5 ), by network depth, on the 2D endoscopy benchmark. Three seeds; per-layer standard deviation across seeds 0.16.
Figure 4: L0 gate allocation on the 2D endoscopy benchmark. (a) Mean open taps per depthwise filter against network depth, at six penalty strengths, with error bars over three seeds. The profile is V-shaped at every λ : the gates retain spatial capacity in the shallow encoder, shed it fastest at up1.conv1 , the first decoder stage, fed directly by the bottleneck, and restore it through the shallow decoder as resolution returns. The ordering is preserved across a 500-fold range of λ and across seeds (mean per-layer standard deviation 0.16 at λ=10−5 , maximum 0.38). Table 6 reports the grouped means at λ=10−5 . (b) Total tap budget against accuracy, by seed. Reducing spatial parameters sevenfold produces no detectable change, while the gap between seeds ( 0.098 Dice between seed means) is nearly twice the largest variation any λ produces within a seed ( 0.056 ).
Variant
Parameters
Dice
Dense U-Net
17,267,393
0.805 ± 0.010
Fixed shift, with padding projection
4,658,780
0.798 ± 0.023
Fixed shift, no projection
1,930,433
0.794 ± 0.003
Learnable depthwise
1,978,268
0.816 ± 0.022
Hybrid
1,954,076
0.811 ± 0.004
Table 7: 2D endoscopy benchmark (Kvasir-SEG [ 15 ] ), Dice on a fixed held-out split, mean ± standard deviation over three random initializations. The split is held fixed across all runs ( split_seed =0 , independent of the run seed) so that variation reflects initialization only, and comparisons against the dense baseline are paired by seed.
Figure 5: Kvasir-SEG segmentations from the hybrid configuration ( up1,up2,up3 frozen, 1,954,076 parameters), on the fixed test split of 100 images ( split_seed =0 ). Rows are the 5th, 50th and 95th percentile of test Dice, ranked, so they are representative rather than selected: 0.98 , 0.89 and 0.26 . Across the split the mean is 0.800 but the median is 0.889 , with a minimum of 0.000 and a maximum of 0.986 , as on MSD Liver, most cases are segmented well and a tail of failures accounts for most of the shortfall in the mean.
Frozen stages
Parameters
Filters frozen
Dice
up1
1,964,444
29%
0.819 ± 0.011
up1,up2
1,957,532
43%
0.810 ± 0.013
up1,up2,up3
1,954,076
51%
0.815 ± 0.008
down4,up1,up2,up3
1,944,860
70%
0.808 ± 0.015
down3,down4,up1,up2,up3
1,937,948
84%
0.814 ± 0.012
Table 8: Freezing configuration sweep on the 2D endoscopy benchmark, five seeds each, fixed split. Configurations are cumulative, beginning at the deepest decoder stage. No configuration differs significantly from any other (one-way ANOVA F=0.61 , p=0.66 ), and all exceed the fully frozen model ( 0.794±0.003 ).
Figure 6: Per-case Dice under five-fold cross-validation over all 131 volumes. (a) tumor Dice for every tumor-bearing volume, ranked. The error is not distributed around the mean: a body of adequate segmentations is followed by 16 volumes ( 14% ) in which the model detects no tumor at all, shown in grey. (b) Liver Dice for the same volumes, on the same evaluation, for contrast, case-level standard deviation 0.032 against 0.334 for tumor.
Figure 7: Per-case tumor Dice against ground-truth lesion volume, log scale. Open circles are the volumes missed entirely. Performance degrades systematically with decreasing burden (Spearman ρ=0.53 , p=5×10−10 , n=118 ): the failures are concentrated in small-lesion cases rather than distributed at random.
Figure 8: Representative segmentations spanning the range in Figure 6 . Each row shows one axial slice at the largest tumor cross-section: CT left, ground truth center, prediction right. Top, volume 100, a large confluent lesion, tumor Dice 0.93 . Middle, volume 122, multiple smaller lesions of which one is recovered, 0.63 . Bottom, volume 55, a 916 -voxel lesion for which the model predicts no tumor at all, 0.00 ; the liver is nonetheless segmented accurately at 0.95 . All panels use a single fixed intensity window spanning the full preprocessed range, [−17,201] HU, so brightness is comparable across the three volumes.
Volume
CNR
Liver coverage
Recall
Precision
Interpretation
39
0.02
61%
0.00
0.05
lesion isodense with parenchyma
104
3.65
1%
0.00
0.19
hypodense lesion excluded from liver
55
5.17
4%
0.00
0.00
hypodense lesion excluded from liver
49
1.07
100%
0.69
0.11
detected, over-segmented
21
1.07
100%
0.42
0.42
genuinely difficult
Table 9: The weakest test volumes, with contrast-to-noise ratio (CNR) between tumor and surrounding parenchyma, and the fraction of ground-truth tumor covered by the predicted liver region.
Dense 3D U-Net
Ours
Ratio
Parameters
12,946,851
536,990
24×
Model file size (float32)
49.4MiB
2.05MiB
24×
Training activation memory
4.58 GB
6.06 GB
0.76×
CPU inference, s/patch
20.6
16.8
1.2×
CPU inference, whole volume
8.5min
8.0min
1.06×
Table 10: Measured consequences of the parameter reduction. Inference timed on a single CPU core at 1283 , forward pass only; activation memory measured by summing module output tensors during a forward pass at 1283 , batch size 1, base width 32. A ratio below one indicates the factorized model is worse on that axis.
Automatic brain tumor segmentation from multi-modal MRI remains challenging because volumetric models often incur substantial computational cost. This paper presents DALight-3D, a compact 3D U-Net variant that combines depthwise separable 3D convolutions, identifier-conditioned normalization, cross-slice attention, and adaptive skip fusion. The method is evaluated on the Medical Segmentation Decathlon Task01 BrainTumour benchmark under matched optimization settings against standard 3D U-Net, Attention U-Net, Residual 3D U-Net, and V-Net baselines. In the reported 50-epoch comparison, DALight-3D achieves a mean Dice of 0.727 with 2.22M parameters, compared with 0.710 Dice and 3.20M parameters for Residual 3D U-Net. Component-wise ablations show consistent performance degradation when SepConv, identifier-conditioned normalization, CSA, or SSFB is removed. These results indicate that DALight-3D offers a favorable accuracy-efficiency trade-off within the present benchmark setting.
Nand Kumar Mishra, Dhruv Mishra, Dr Manu Pratap Singh
Department of Computer Science, Dr. Bhimrao Ambedkar University, Agra, India. · Department of Computer Science and Engineering, Shiv Nadar University, Greater Noida, India.
We propose parameter-efficient SSM-based U-Net architectures for 3D medical image segmentation. Convolutional U-Nets afford O(n) local mixing per layer but lack explicit global context; transformers provide global reasoning at O(n^2) cost in sequence length n. State-space models (SSMs), such as Mamba, offer O(n) global propagation per block. Yet, existing medical SSM segmenters rely on fixed scan patterns and large parameter budgets. Dynamic Adaptive Scan (DAS), which learns data-dependent reordering before selective scan, has not been applied to medical imaging or extended to 3D volumes. We propose DAMamba-UNet3D, a hybrid encoder-decoder that integrates tri-plane 3D-DAS blocks at encoder stages E2-E4 while retaining convolutions elsewhere (~5.3M parameters). On BraTS 2020 five-fold cross-validation, DAMamba-UNet3D achieves mean Dice 0.815+/-0.013 (full-volume per-case evaluation) at ~13x lower parameter cost than SegMamba (0.824+-0.014, ~70M). At comparable scale, DAMamba-L (~70M), a wide DAS-native variant with encoder-only DAMamba and a convolutional bottleneck, reaches 0.829+-0.012, surpassing retrained SegMamba by 0.5pt. Component ablations show that encoder-only DAS placement is critical as bottleneck and decoder SSM blocks lower Dice. Together, the results suggest that learned tri-plane DAS in a hybrid U-Net is competitive with, and under our large-scale design may improve upon, SegMamba's fixed Tri-orientated Mamba (ToM) scanning on BraTS 2020. Code: https://github.com/marafathussain/DAMamba-UNet3D.
Mohammad Arafat Hussain, Ellen Grant, Yangming Ou
Boston Children’s Hospital, Harvard Medical School
Lightweight 3D medical image segmentation remains constrained by a fundamental *efficiency / robustness conflict''*, particularly when processing complex anatomical structures and heterogeneous modalities. In this paper, we study how to redesign the framework based on the characteristics of high-dimensional 3D images, and explore data synergy to overcome the fragile representation of lightweight methods. Our approach, VeloxSeg, begins with a deployable and extensible dual-stream CNN-Transformer architecture composed of Paired Window Attention (PWA) and Johnson-Lindenstrauss lemma-guided convolution (JLC). For each 3D image, we invoke a glance-and-focus'' principle, where PWA rapidly retrieves multi-scale information, and JLC ensures robust local feature extraction with minimal parameters, significantly enhancing the model's ability to operate with low computational budget. Followed by an extension of the dual-stream architecture that incorporates modal interaction into the multi-scale image-retrieval process, VeloxSeg efficiently models heterogeneous modalities. Finally, Spatially Decoupled Knowledge Transfer (SDKT) via Gram matrices injects the texture prior extracted by a self-supervised network into the segmentation network, yielding stronger representations than baselines at no extra inference cost. Experimental results on multimodal benchmarks show that VeloxSeg achieves a 26% Dice improvement, alongside increasing GPU throughput by 11×, CPU by 48×, and reducing training peak GPU memory usage by 1/20, inference by 1/24. Code is available at https://github.com/JinPLu/VeloxSeg.
Jinpeng Lu, Linghan Cai, Yinda Chen +4
University of Science and Technology of China · Dresden University of Technology · Harbin Institute of Technology, Shenzhen