Multiscale Bottleneck Block (M2B)

• Multiscale Bottleneck Block (M2B) is a modular decoder component that integrates parallel depthwise convolutions, channel shuffle, and residual aggregation for efficient 3D segmentation.
• It refines decoder-stage features by fusing multiscale contexts using 3×3×3, 5×5×5, and 7×7×7 convolutions, balancing detailed and coarse information with low computational overhead.
• Empirical evaluations on Synapse and ACDC datasets show improved Dice scores with minimal increases in parameters and FLOPs compared to standard 3D bottleneck designs.

The Multiscale Bottleneck Block (M²B) is a modular architectural component introduced in the decoder of the MLRU++ (Multiscale Lightweight Residual UNETR++) network to improve 3D medical image segmentation by integrating efficient multiscale contextual refinement with low computational overhead. M²B adapts the MobileNetV2-style inverted residual bottleneck for 3D volumetric contexts and specifically targets decoder-stage feature reconstruction, providing parallel multiscale depthwise convolutions supplemented by channel shuffle and residual aggregation. This design enables effective multi-resolution semantic fusion while maintaining a lightweight profile, addressing the high computational burden often associated with hybrid CNN-Transformer segmentation frameworks (Yadav et al., 22 Jul 2025).

1. Architectural Specification of M²B

In MLRU++, the M²B is positioned after upsampling and skip-fusion within each decoder stage. The input tensor $$x \in \mathbb{R}^{B \times C \times H \times W \times D}$$ undergoes a sequence of operations:

• Pointwise Expansion: $1\times1\times1$ Conv3D, expanding channels from $C$ to $tC$ ($t$: expansion ratio, not explicitly reported).
• BatchNorm and Channel Shuffle: Promotes inter-channel information flow.
• Parallel Multiscale Depthwise Convolutions: Three explicit branches applying depthwise 3D convolutions with kernels $3\times3\times3$, $5\times5\times5$, and $7\times7\times7$ (all with groups $= tC$), each followed by BN and ReLU6.
• Branch Aggregation: Outputs from all three branches are summed elementwise, preserving channel count.
• Residual Addition: The aggregated features are added to the expanded representation (internal residual).
• Nonlinearity and Projection: ReLU6 (placement partially ambiguous in the text) and a $1\times1\times1$ Conv3D projecting back to $1\times1\times1$0 channels, followed by BN.

This block design preserves spatial resolution across branches and avoids explicit feature pyramids or concatenation, focusing instead on receptive-field diversity within the same resolution.

Schematic Workflow

$tC$6

2. Mathematical Formulation and Functional Role

Expansion:

$1\times1\times1$1

Parallel Depthwise Convolutions:

For $1\times1\times1$2,

$1\times1\times1$3

Branch Fusion and Residual:

$1\times1\times1$4

No explicit whole-block residual (input–output skip) is defined in the source, though practical implementation might conditionally include it if $1\times1\times1$5 and spatial sizes match.

The block functions as a decoder-side multiscale context refiner, providing direct access to short- and long-range semantic evidence across organ/tissue scales while maintaining low model complexity typical of depthwise MobileNet-style designs.

3. Computational Complexity and Comparative Analysis

M²B remains computationally lightweight relative to full 3D convolutions by exploiting depthwise separability and summation (rather than concatenation) in branch fusion. The total parameter count (ignoring batchnorm for brevity) is:

$1\times1\times1$6

where the $1\times1\times1$7 term aggregates the three depthwise branches with $1\times1\times1$8.

For comparison, a standard 3D bottleneck block (single full $1\times1\times1$9 Conv) would require:

$C$0

Thus, M²B trades the quadratic $C$1 cost of dense convolution for a linear-in-$C$2 depthwise multibranch complexity, yielding substantial reductions when $C$3 is moderate-to-large and $C$4 small.

The FLOPs are similarly dominated by the sum of pointwise and depthwise operations:

$C$5

Channel shuffle and nonlinearity induce negligible additional overhead.

4. Empirical Evidence and Quantitative Impact

Direct ablations on Synapse and ACDC datasets establish M²B’s empirical benefit:

Dataset	Variant Comparison	Mean DSC Gain	Params Increase	FLOPs Increase
Synapse	LCBAM-only encoder vs (LCBAM+M²B)	+1.19	+1.17M	+1.25G
Synapse	LCBAM encoder+decoder vs (LCBAM+M²B)	+1.49	+1.49M	+2.43G
ACDC	MLRU++ w/o M²B vs full MLRU++	+0.41	–	–

For Synapse, the Dice coefficient improves by approximately $C$6 to $C$7 points with modest increases in parameter count and FLOPs. For ACDC, the gain is $C$8 mean Dice, with consistent improvements across cardiac subregions (RV, Myocardium, LV). No isolated M²B-only ablation is available for BTCV or Lung; their full-model Dice scores with M²B present are 85.20 and 81.12, respectively.

Inference time and memory footprint attributable specifically to M²B are not reported; efficiency is addressed primarily via aggregate parameter and FLOP counts at model level (Yadav et al., 22 Jul 2025).

5. Multiscale Mechanism and Decoder Integration

M²B’s multiscale property is intra-block, achieved by parallel branches with $C$9, $tC$0, and $tC$1 depthwise kernels acting on the same expanded input tensor. This design captures:

• Fine-grained local details ($tC$2)
• Medium-range context ($tC$3)
• Coarse, large-scale spatial context ($tC$4)

All branches maintain spatial size via padding, enabling direct elementwise summation. M²B operates after each decoder upsampling and skip fusion, refining stage-aligned features rather than performing resolution alignment itself. Summation rather than concatenation preserves channel dimensionality and promotes lightweight computation, consistent with MobileNet design heuristics.

6. Design Variations, Hyperparameters, and Reproducibility Considerations

The specification of M²B in the literature is precise in high-level operation but omits certain implementation-critical details, including:

• Expansion ratio $tC$5 (controls hidden width)
• Exact decoder stage channel mapping
• Placement of BN/ReLU6 around expansion/projection layers
• Scope of the residual connection (internal vs whole-block)

Consequently, reference implementations must choose these values consistently with common MobileNet/UNetR++ conventions to replicate reported efficiency and accuracy. No in-depth ablation is presented for kernel multiplicity, expansion ratio, shuffle inclusion, or branch aggregation variant (only presence/absence of M²B as a whole is examined), and thus architectural variants outside the described form are not empirically characterized (Yadav et al., 22 Jul 2025).

7. Contextual Significance and Performance Interpretation

M²B’s role as an intra-stage multiscale refinement unit is substantiated by gains in Dice coefficient, particularly on boundary-rich and structurally small anatomical regions. The block enables decoder representations to simultaneously access diverse spatial contexts at low computational cost, which is empirically advantageous for 3D medical segmentation tasks suffering from high anatomical variability and limited memory budgets. Application within MLRU++ leads to state-of-the-art or competitive performance on multiple 3D benchmarks with favorable parameter and FLOP profiles relative to dense convolutional or hybrid CNN-Transformer alternatives.

While empirical impact and design rationale are robustly motivated for Synapse and ACDC, results on BTCV and Lung are reported only at the aggregate level, and reproducibility requires attention to unspecified hyperparameters and internal block conventions.

Reference: (Yadav et al., 22 Jul 2025)