Semantic Learning for Molecular Communication in Internet of Bio-Nano Things

Hanlin Cai, Özgür B. AkanarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces a novel end-to-end semantic learning framework for molecular communication (MC) in the Internet of Bio-Nano Things (IoBNT). By prioritizing task-relevant information and modeling the inherent stochasticity of MC channels, the framework significantly improves diagnostic accuracy in resource-constrained environments.

What Problem Does This Paper Address?

Molecular communication (MC) holds immense promise for applications within the Internet of Bio-Nano Things (IoBNT), such as disease diagnostics and targeted drug delivery, due to its biocompatibility and energy efficiency. However, MC channels are plagued by inherent limitations like low data rates, inter-symbol interference (ISI), and susceptibility to noise. These challenges severely restrict the complexity of data that can be reliably transmitted, hindering advanced applications. Traditional communication methods struggle to overcome these issues, especially in resource-constrained scenarios where every molecule counts. This paper tackles the problem of efficiently and reliably transmitting complex information, specifically for diagnostic tasks, over these challenging molecular channels.

Methods

The core of this work is an end-to-end semantic learning framework built upon a deep encoder-decoder architecture. Here's a breakdown of the key components:

  • Semantic Feature Extraction: An encoder, implemented as a Convolutional Neural Network (CNN) followed by a linear layer, extracts high-level, task-relevant semantic features from the input data (biomedical images in this case). This process aims to distill the most crucial information while discarding redundancy.
  • Probabilistic Quantization: A novel quantization function maps these semantic features into a normalized vector of channel input symbols. This step is crucial for translating abstract semantic information into physically transmittable molecular parameters, ensuring that the transmission adheres to physical constraints on molecular release.
  • Probabilistic Channel Network: To address the non-differentiable and stochastic nature of molecular propagation, a probabilistic channel network is introduced. This network models the conditional probability distribution of received symbols given transmitted symbols as a mixture of Gaussian distributions. This allows for gradient-based optimization of the entire system.
  • Task-Specific Decoding: A decoder, a fully connected neural network, takes the received channel symbols and maps them to the task-specific output (e.g., probability distribution for diagnostic classification).
  • End-to-End Training: The entire framework, from encoder to decoder, is trained jointly. This allows the system to optimize for the specific diagnostic task while simultaneously adapting to the complexities of the molecular channel. The cross-entropy loss function is used to minimize the difference between ground truth and predicted outputs.

Datasets

The framework is evaluated using the Kvasir dataset [12]. This dataset comprises 8,000 high-resolution gastrointestinal images categorized into eight clinically significant classes, representing normal findings and various pathological conditions. The images are preprocessed to 128x128 pixels, preserving RGB channels for diagnostic relevance.

Key Insights and Ideas

  • Semantic Prioritization is Key: The paper strongly advocates for moving beyond bit-level transmission in MC. By focusing on semantic information, the system can achieve higher performance even with limited molecular resources.
  • Modeling Channel Stochasticity is Crucial: The introduction of a probabilistic channel network is a significant step. It allows for end-to-end learning by providing a differentiable approximation of the complex, noisy, and ISI-prone molecular channel.
  • Transferability and Adaptability: The semantic approach, coupled with the probabilistic channel model, enhances the framework's ability to adapt to different channel conditions and diverse IoBNT tasks.
  • ISI Can Be Beneficial (in a way): Counter-intuitively, the paper notes that networks trained with lower maximum molecule release capacities (nm) exhibit better learning efficiency. This is because the increased impact of ISI and channel degradation forces the model to learn more robust propagation features.

Key Results

  • The proposed semantic learning framework achieves at least a 25% performance gain in diagnostic accuracy compared to conventional JPEG compression with LDPC coding methods under resource-constrained conditions.
  • The framework demonstrates superior efficiency and robustness over traditional methods in diagnostic classification tasks.
  • The probabilistic channel network accurately replicates the physical characteristics of molecular propagation, capturing pronounced ISI in low flow velocity scenarios and accelerated molecular clearance in high flow velocity scenarios, as visualized in Figure 2.

Limitations and Future Work

  • Limitations: The current framework is designed for Single-Input Single-Output (SISO) molecular communication systems. The paper acknowledges that the performance gains are observed under specific resource-constrained conditions.
  • Future Work: The authors propose extending the framework to Multiple-Input Multiple-Output (MIMO) scenarios, exploring more sophisticated adaptive semantic extraction techniques, and further enhancing the overall efficiency for real-world IoBNT deployments.

Summary updated

BibTeX
@misc{cai2025semantic,
  author = {Hanlin Cai and Özgür B. Akan},
  title = {{Semantic Learning for Molecular Communication in Internet of Bio-Nano Things}},
  year = {2025},
  eprint = {2502.08426},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.08426},
}
Figure