Lung-SRAD: Spectral-Aware Regularized Audio DASS with Dual-Axis Patch-Mix Contrastive Learning for Respiratory Sound Classification

Technical Summary: Lung-SRAD

Problem Statement
Respiratory Sound Classification (RSC) is critical for diagnosing conditions like pneumonia, COPD, and asthma, where abnormal sounds such as crackles and wheezes manifest as short-duration, localized spectro-temporal structures. While recent high-performing RSC methods rely on Transformer-based architectures like the Audio Spectrogram Transformer (AST), the authors identify two primary limitations:

1. Spectral Bias: Theoretical analysis suggests that softmax-based self-attention in Transformers behaves as a low-pass filter in the feature domain. This progressively suppresses high-frequency inter-token variations, potentially reducing sensitivity to the localized, rapid spatial variations characteristic of abnormal respiratory events.
2. Computational Complexity: The quadratic time and memory complexity of self-attention makes Transformers computationally expensive for long audio sequences.

Methodology: Lung-SRAD
The paper proposes Lung-SRAD (Spectral-Aware Regularized Audio DASS), a framework built upon State Space Models (SSMs) to address these limitations. The methodology consists of three core components:

1. SSM Backbone (DASS):
   The authors utilize the Distilled Audio State Space model (DASS), initialized with weights distilled from an ensemble of Transformer teachers (AST and HTS-AT) on the AudioSet. Unlike Transformers, DASS employs a hierarchical architecture with 2D Selective State Space Scanning (SS2D). This allows the model to aggregate features from the full time-frequency map via spatial pooling rather than relying on a single CLS token, maintaining linear complexity while preserving distributed representations.

2. Spectral-Aware Layer Regularization:
   Through an analysis of spectral response curves, the authors observed that while DASS preserves mid-to-high spatial-frequency components better than AST, certain intermediate layers (specifically in Stage 2) exhibit excessive high-frequency responses that may lead to overfitting on fine-grained local patterns. To mitigate this, they apply Gaussian convolution selectively to these layers. This regularization attenuates dominant spectral peaks while preserving the overall harmonic structure, aiming to improve specificity without sacrificing sensitivity.

3. Dual-Axis Patch-Mix Contrastive Learning:
   Standard Patch-Mix contrastive learning, designed for AST, treats spectrogram patches as an unordered set. However, SSMs like VMamba process inputs as structured 2D grids with directional state transitions. Naive patch mixing disrupts these sequential state evolutions. To align with the SSM architecture, the authors propose Dual-Axis Patch-Mix:

   • Temporal Patch-Mix: Replaces consecutive temporal segments while preserving frequency-wise state continuity.
   • Frequency Patch-Mix: Replaces consecutive frequency bands while preserving temporal continuity.
   • Asymmetric Gradient Strategy: The model uses a stop-gradient operation on the mixed samples to encourage invariance to these axis-aligned perturbations without collapsing the backbone's structured state-space dynamics.

Key Contributions

• First Application of DASS for RSC: The paper presents the first use of a distilled State Space Model for respiratory sound classification, achieving a baseline score of 61.06% on the ICBHI dataset via simple fine-tuning.
• Spectral Analysis and Regularization: The authors provide an empirical analysis of intermediate representations in SSMs versus Transformers, demonstrating DASS's superior preservation of mid-to-high spatial-frequency components. They introduce a targeted Gaussian regularization strategy to balance these high-frequency responses.
• Architecture-Aligned Contrastive Learning: They propose a Dual-Axis Patch-Mix supervised contrastive learning strategy specifically tailored to the multi-directional scanning properties of SSMs, avoiding the structural disruption caused by standard patch-mixing techniques.

Experimental Results
Experiments were conducted on the ICBHI benchmark using the official 60% training and 40% test split (patient-independent).

• 4-Class Evaluation: The final Lung-SRAD model achieved an ICBHI Score of 64.48%, outperforming the AST baseline (62.37% with Patch-Mix CL) by approximately 2.1 percentage points and the standard AST fine-tuning baseline by 5 percentage points. This improvement was driven by a significant increase in Specificity (Sp) from 74.68% to 79.53%, while maintaining comparable Sensitivity (Se).
• 2-Class Evaluation: In a binary classification setting (Normal vs. Abnormal), the model achieved a score of 72.57%, exceeding the previous state-of-the-art reported in the paper (72.08%).
• Ablation Studies: The results confirmed that spectral-aware regularization improved specificity, and the Dual-Axis Patch-Mix strategy provided superior regularization compared to single-axis mixing or AST-style mixing.

Significance and Claims
The paper claims that Lung-SRAD demonstrates the viability of State Space Models as an efficient and effective alternative to Transformers for RSC. By leveraging the inherent ability of SSMs to preserve localized spectro-temporal patterns and introducing architectural-specific regularization and contrastive learning, the method achieves a favorable balance between sensitivity and specificity. The authors emphasize that their approach addresses the specific spectral limitations of attention-based models in the context of detecting sparse, abnormal respiratory events, offering a computationally efficient pathway for clinical audio analysis.