SCDM: Spatial-Contextual Disentanglement Mamba via Differential Inference for Efficient Image Classification

1 Introduction

• CNNs such as ResNet leverage local inductive biases and have set competitive benchmarks, but fixed kernels limit long-range global context needed for subtle pathological patterns.Medical image classification supports early detection of pathologies.Source: §1 · pp. 1–2
• SSMs such as Mamba offer linear-complexity long-range context modeling as an alternative to the quadratic computational cost of Vision Transformer self-attention [7] [8].ViTs and Swin Transformers use self-attention for global spatial modeling, while CNNs' fixed kernels limit long-range context capture.Source: §1 · p. 2
• SCDM addresses entanglement between pathological features and visually similar anatomy with a Positive Branch for discriminative features and a Negative Branch that models and suppresses normal anatomical mimics.Its similarity-driven repulsion and differential inference are designed to isolate pathology-relevant signals without increasing model complexity or parameters.Source: §1 · p. 2

2 Preliminaries

• SCDM extends the standard VMamba backbone with a differential dual-branch extension at the SS2D level, without changing the overall backbone structure.The extension enables explicit supervision and disentanglement of hidden-state representations.Source: §2 · p. 2
• SSMs map an input sequence x(t) to an output y(t) through a hidden state h(t), and Mamba's S6 mechanism uses input-dependent discretization.The sequence-to-state formulation is cited to [5].Source: Selective State Space Modeling · p. 2
• SS2D extends SSMs to 2D by scanning feature maps along multiple spatial routes and fusing outputs for global receptive fields with linear complexity.The routes use cross-directional scans.Source: Selective State Space Modeling · p. 2
• A VSS block combines an SS2D spatial-modeling branch with a feed-forward channel-transformation branch in a residual framework.Stacked VSS blocks form the VMamba backbone.Source: Selective State Space Modeling · p. 2

3.1 Overview

• SCDM uses an asymmetric dual-branch design under image-level supervision, with cooperative branch addition during training and competitive differential readout during inference.The Negative Branch suppresses normal anatomical structures such as ribs while the Positive Branch focuses on lesions.Source: §3.1 · p. 3

3.2 Dual-Branch State Space Formulation

• The Negative Branch models normal anatomical context, while the Positive Branch detects pathological deviations; the Positive Branch input is dynamically modulated by a repulsion gate.This formulation aims to learn complementary representations of anatomy and pathology.Source: §3.2 · pp. 3–4

3.3 Repulsion Gating Mechanism

• The repulsion gate compares the current input with aggregated Negative Branch memory and uses that similarity to suppress features already explained by the Negative Branch.A warm-up phase disables the gate so both branches can establish foundational representations before similarity-driven competition begins.Source: §3.3 · pp. 4–5

3.4 Integration into Vision Selective Scan (VSSc) Blocks

• Intermediate VSSc layers merge branch outputs, but the final VSSc block keeps them separate; pooled representations feed a classifier based on their competitive difference.Independent feed-forward networks process the final outputs, and the competitive formulation is intended to reduce overlap between contextual and pathology-sensitive features.Source: §3.4 · p. 5
• Training adds branch logits for cooperative signal accumulation, while competitive separation and diseased-sample orthogonality losses promote branch specialization.An activity-preservation loss enforces a minimum Positive Branch activation norm to prevent zero-activation collapse and support stable gradient propagation.Source: §3.4 · pp. 5–6

4 Experiments

• The evaluation uses the RSNA pneumonia dataset, an 80/20 patient-wise split, and 128×128 images, with a shallower SCDM stage configuration than VMamba-B.The study also applies CLAHE preprocessing, augmentations, and a WeightedRandomSampler for class imbalance.Source: §4 · p. 6

4.1 Comparison with Accuracy Metrics

• SCDM achieves competitive AUC with substantially lower recall than ResNet-101, outperforms Swin-B, ViT-Base, and Vmamba-B, and uses 29.4M parameters and 1.44 GFLOPs.Table 1 compares performance and efficiency across CNN-, transformer-, and state-space models.Source: Table 1 · p. 6, §4.1 · p. 6
• SCDM maintains strong specificity with fewer parameters and FLOPs; Fig. 2 compares computational cost and model specifications.The text attributes stronger true-negative discrimination to the dual-memory mechanism without increased model capacity.Source: Figure 2 · p. 7, §4.1 · p. 7

4.2 Representation-Level Analysis and Localization Behavior

• Grad-CAM analysis reports more precise SCDM activation patterns than VMamba’s diffuse maps, with differential inference refining the Positive Branch’s focus.Fig. 3 compares localization patterns between SCDM and representative baselines.Source: §4.2 · pp. 7–8, Figure 3 · p. 7

4.3 Component and Training Analysis

• With differential readout fixed, cooperative BCE(P + N) training supports anatomical-context scaffolding, whereas competitive BCE(P − N) training limits robust anchors and overall Recall.Table 2 compares these training objectives; Fig. 4 is described as showing more selective localization for the asymmetric approach than standard fused representations.Source: §4.3 · p. 8, Table 2 · p. 8, Figure 4 · p. 8
• The Repulsion Gate drives sensitivity: penalizing anatomical overlap shifts the operating point toward high Recall, with a 5% gain and marginally lower precision.The text says this prioritizes minimizing false negatives in clinical screening, while the integrated SCDM offers the most robust balance.Source: §4.3 · p. 8

• SCDM is reported to improve sensitivity and localization by suppressing hard negatives while requiring fewer parameters and lower computational cost.Source: §5 · p. 9
• Dual-branch differential inference is used to disentangle focal pathology from anatomical distractors.Source: §5 · p. 9
• Future work targets multi-class, high-resolution scaling.Source: §5 · p. 9