Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving Liver Disease Diagnosis with SNNDeep: A Custom Spiking Neural Network Using Diverse Learning Algorithms

Zofia Rudnicka, Janusz Szczepanski, Agnieszka Pregowska

Abstract

Purpose: Spiking neural networks (SNNs) have recently gained attention as energy-efficient, biologically plausible alternatives to conventional deep learning models. Their application in high-stakes biomedical imaging remains almost entirely unexplored. Methods: This study introduces SNNDeep, the first tailored SNN specifically optimized for binary classification of liver health status from computed tomography (CT) features. To ensure clinical relevance and broad generalizability, the model was developed and evaluated using the Task03\Liver dataset from the Medical Segmentation Decathlon (MSD), a standardized benchmark widely used for assessing performance across diverse medical imaging tasks. We benchmark three fundamentally different learning algorithms, namely Surrogate Gradient Learning, the Tempotron rule, and Bio-Inspired Active Learning across three architectural variants: a fully customized low-level model built from scratch, and two implementations using leading SNN frameworks, i.e., snnTorch and SpikingJelly. Hyperparameter optimization was performed using Optuna. Results: Our results demonstrate that the custom-built SNNDeep consistently outperforms framework-based implementations, achieving a maximum validation accuracy of 98.35%, superior adaptability across learning rules, and significantly reduced training overhead. Conclusion:This study provides the first empirical evidence that low-level, highly tunable SNNs can surpass standard frameworks in medical imaging, especially in data-limited, temporally constrained diagnostic settings, thereby opening a new pathway for neuro-inspired AI in precision medicine.

Improving Liver Disease Diagnosis with SNNDeep: A Custom Spiking Neural Network Using Diverse Learning Algorithms

Abstract

Purpose: Spiking neural networks (SNNs) have recently gained attention as energy-efficient, biologically plausible alternatives to conventional deep learning models. Their application in high-stakes biomedical imaging remains almost entirely unexplored. Methods: This study introduces SNNDeep, the first tailored SNN specifically optimized for binary classification of liver health status from computed tomography (CT) features. To ensure clinical relevance and broad generalizability, the model was developed and evaluated using the Task03\Liver dataset from the Medical Segmentation Decathlon (MSD), a standardized benchmark widely used for assessing performance across diverse medical imaging tasks. We benchmark three fundamentally different learning algorithms, namely Surrogate Gradient Learning, the Tempotron rule, and Bio-Inspired Active Learning across three architectural variants: a fully customized low-level model built from scratch, and two implementations using leading SNN frameworks, i.e., snnTorch and SpikingJelly. Hyperparameter optimization was performed using Optuna. Results: Our results demonstrate that the custom-built SNNDeep consistently outperforms framework-based implementations, achieving a maximum validation accuracy of 98.35%, superior adaptability across learning rules, and significantly reduced training overhead. Conclusion:This study provides the first empirical evidence that low-level, highly tunable SNNs can surpass standard frameworks in medical imaging, especially in data-limited, temporally constrained diagnostic settings, thereby opening a new pathway for neuro-inspired AI in precision medicine.

Paper Structure

This paper contains 22 sections, 8 equations, 4 figures, 19 tables.

Table of Contents

  1. Introduction
  2. Materials and Methods
  3. Datasets and preprocessing
  4. Dataset 1: Task03_Liver
  5. Dataset 2: Task03_CECT
  6. Patient-level splitting and evaluation protocol
  7. Temporal spike encoding
  8. Neuron model
  9. Network architectures
  10. Surrogate Gradient Learning
  11. Tempotron learning
  12. Patient-level aggregation and metrics
  13. Statistical analysis
  14. Computational efficiency
  15. Implementation details and reproducibility
  16. ...and 7 more sections

Figures (4)

  • Figure 1: Architecture of the convolutional spiking neural network. Each preprocessed CT slice is represented as a $96\times96\times15$ tensor obtained from 2.5D axial context and three HU windows. The input is processed by three convolutional leaky integrate-and-fire blocks and mapped to a single logit through global average pooling and a linear classifier.
  • Figure 2: Architecture of the ConvSNN-MIL model. Each patient is represented as a bag of axial slices. Every slice is processed by a shared convolutional spiking encoder, yielding slice-level feature vectors. These features are then aggregated by attention-based multiple-instance learning pooling to obtain a patient-level representation, which is mapped to the final lesion-presence probability.
  • Figure 3: Seed-wise patient-level PR-AUC across five independent leakage-free patient-level splits for the two evaluated CT datasets. Panel (a) shows results for Task03_Liver, and panel (b) for Task03_CECT. The second dataset exhibited lower absolute PR-AUC and greater variability across seeds, indicating stronger sensitivity to the particular patient-level partition.
  • Figure 4: Distribution of patient-level performance across five independent random patient-level splits for both evaluated datasets. The top row shows Task03_Liver, and the bottom row shows Task03_CECT. Across the main metrics, Task03_Liver exhibited higher and more stable performance, whereas Task03_CECT showed a broader performance spread and greater dataset-dependent variability.