Hierarchical DenseNet Swint Architecture

• The paper demonstrates that integrating a modified DenseNet-201 stem with a Swin Transformer branch significantly enhances brain MRI classification accuracy (98.50%) by capturing both fine-grained textures and global morphology.
• The methodology employs dual residual connections and deep feature extraction modules to fuse features across scales, effectively preserving and aligning local and global diagnostic cues.
• The architecture’s hierarchical integration of local-to-global features reduces false negatives and proves robust in distinguishing diverse tumor classes with complex morphological traits.

The Hierarchical DenseNet Swint Architecture is a convolutional–transformer hybrid designed to achieve superior performance in large-scale brain MRI classification by integrating fine-grained texture modeling with robust global context reasoning. Architecturally, it fuses a modified DenseNet-201 convolutional stem with a Swin Transformerₜ branch, employing deep feature extraction blocks and dual residual pathways. This hybridization enables discriminative, hierarchical feature fusion—critical for maximizing sensitivity and specificity across tumor classes with diverse morphological and textural traits (Shah et al., 26 Jan 2026).

1. Overall Network Topology

Input to the system is a single MR slice of shape $224 \times 224 \times 3$. The architecture bifurcates into two parallel stems:

• DenseNet-201 Stem ("Stem CNN"): Customized for MRI spatial statistics, it focuses on structured local feature extraction through a progression of convolution, pooling, and densely connected composite blocks. The output feature map dimensions and channels progress through:
  • Conv1: $7 \times 7$ kernel, $64$ channels, stride $2 \to 112 \times 112 \times 64$
  • MaxPool: $3 \times 3$ kernel, stride $2 \to 56 \times 56 \times 64$
  • Series of DenseBlocks: escalating depth and channel count, interleaved with transitional convolutions and average pooling, culminating in a global average pooled tensor $F_d \in \mathbb{R}^{1 \times 1 \times 1920}$.
• Swin Transformerₜ Stem ("Global-context Branch"): Implements hierarchical self-attention for global morphological modeling.
  • PatchEmbedding partitions the input into $4 \times 4$ patches, each flattened and linearly projected to $d=96$ dimensions, creating $56 \times 56$ token grids.
  • Four Transformer stages with alternations of SwinBlocks and PatchMerging sequentially reduce spatial dimensions and increase channel dimensionality, concluding with a global average pooled feature $7 \times 7$0.

This dual-stem approach ensures simultaneous encoding of local micro-structure (e.g., irregular edges and texture variation) and holistic anatomical context (e.g., tumor positioning, mass effect).

2. Deep Feature Extraction (DFE) and Dual Residual (DR) Connection Mechanisms

Deep Feature Extraction (DFE) modules are interspersed within the DenseNet stem to bolster feature propagation and retention. Each DFE block applies a pair of convolution–BN–ReLU stacks with an intra-block skip connection: $7 \times 7$1

$7 \times 7$2

where $7 \times 7$3 denotes 2D convolution and $7 \times 7$4 is ReLU activation.

The Dual Residual (DR) route orchestrates feature fusion at both the input and output of the Swin Transformer:

• First Residual: Projects $7 \times 7$5 into the Swin token space, broadcasts across tokens, and concatenates with patch-embedded Swin input $7 \times 7$6 along the feature axis:

$7 \times 7$7

where $7 \times 7$8.

• Second Residual: Fuses Swin output $7 \times 7$9 with a projection of $64$0 via an additive skip-connection:

$64$1

The classifier operates on the flattened $64$2:

$64$3

This dual-residual formulation preserves both shallow and deep representations, mitigating feature decay and facilitating effective cross-branch conditioning.

3. Swin Transformer Staging and Multi-Scale Self-Attention

The Swin Transformerₜ branch utilizes windowed self-attention to balance computational efficiency with contextual reach:

• Patch Embedding: Formalized as

$64$4

• Window Partitioning & Shift: Feature maps $64$5 are split into $64$6 non-overlapping windows. Shifted windows are generated by a spatial shift of $64$7, then repartitioned.
• Window-based Multi-Head Self-Attention: Each window receives standard transformer self-attention

$64$8

with $64$9 projected from window tokens and $2 \to 112 \times 112 \times 64$0 a pre-computed relative position bias.

This staging enables the model to encode both locally concentrated and spatially distant relationships, tailored by shifted windowing to reduce information bottlenecking across window boundaries.

4. Feature-Dimension Alignment and Fusion Strategies

For consistent feature merging, both branches undergo channel and spatial alignment before fusion:

• Projections linearly map $2 \to 112 \times 112 \times 64$1 and $2 \to 112 \times 112 \times 64$2 to a common latent:

$2 \to 112 \times 112 \times 64$3

• If sequence dimensions misalign, spatial upsampling or downsampling is performed to match $2 \to 112 \times 112 \times 64$4 with $2 \to 112 \times 112 \times 64$5.
• Final fusion is executed as

$2 \to 112 \times 112 \times 64$6

with channelwise concatenation $2 \to 112 \times 112 \times 64$7 and nonlinearity $2 \to 112 \times 112 \times 64$8 as ReLU or GELU.

This systematic alignment is necessary for robust, information-preserving combination prior to the softmax classifier.

5. Hierarchical Integration: Local-to-Global Feature Synergy

The hierarchical design underpins both multi-scale and multi-level integration:

• Early DenseNet layers (DenseBlocks 1 & 2) capture minute edge and textural cues.
• Deeper Swin stages (Stages 3 & 4) acquire coarse-scale global morphology and cross-cutting context.

At multiple resolution levels $2 \to 112 \times 112 \times 64$9, local DenseNet features $3 \times 3$0 are projected and optionally concatenated or summed with Swin branch activations. For this implementation, the two-residual scheme fuses all DenseNet features at the Swin input, with late fusion after the last Swin block. This bilateral exchange supports conditioning of global morphology on precise, spatially-localized cues, and vice versa.

The hierarchical fusion strategy, symbolically: $3 \times 3$1 enables context-aware discrimination—vital for invariance to tumor size, texture, and anatomical idiosyncrasies.

6. Empirical Performance and Diagnostic Significance

Evaluation on a rigorously curated MRI dataset—40,260 images, four classes—resulted in a test set accuracy and recall of $3 \times 3$2. This architecture outperformed standalone CNNs, Vision Transformers, and simple hybrids (Shah et al., 26 Jan 2026). Notably, by learning both irregular and diffuse glioma traits (via the Boosted Feature Space setup) and the well-defined mass/location of meningioma or pituitary tumors (via the hierarchical DFE+DR paradigm), the approach demonstrably suppresses false negatives across all tumor entities. The clinical implication is high sensitivity for complex tumor phenotypes and reliability in identifying both textural and morphological aberrations.

Architecture Component	Primary Role	Output Dimensionality
DenseNet-201 Stem	Local convolutional texture extraction	$3 \times 3$3
Swin Transformerₜ Stem	Long-range dependency, morphology modeling	$3 \times 3$4
Dual Residual Fusion	Hierarchical feature integration	$3 \times 3$5

This suggests the utility of deep, hierarchical architectures with dual residual channels for high-stakes, morphology-variant medical image diagnostics. A plausible implication is the extensibility of this framework to other organ systems or modalities demanding simultaneous fine and global feature modeling.