💻 computer science

Reliable Brain Tumor Segmentation Based on Spiking Neural Networks with Efficient Training

This paper proposes a reliable and energy-efficient 3D brain tumor segmentation framework using a multi-view spiking neural network ensemble trained with Forward Propagation Through Time (FPTT), which achieves competitive accuracy and well-calibrated uncertainty on BraTS datasets while reducing computational costs by 87%.

Original authors: Aurora Pia Ghiardelli, Guangzhi Tang, Tao Sun

Published 2026-01-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Aurora Pia Ghiardelli, Guangzhi Tang, Tao Sun

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a hidden treasure (a brain tumor) inside a giant, complex 3D block of jelly (a brain MRI scan). Usually, doctors use powerful, energy-hungry computers to slice through this jelly and mark exactly where the treasure is. But these computers are like heavy, gas-guzzling trucks—they need a lot of power and can't easily fit into small, portable devices.

This paper introduces a new, lightweight "electric scooter" approach to solving this problem. Here is how it works, broken down into simple concepts:

1. The Brainy "Spiking" Neurons (The Electric Scooter)

Instead of using standard computer brains (Deep Neural Networks) that constantly buzz with electricity, the authors built a system using Spiking Neural Networks (SNNs).

The Analogy: Think of a standard computer brain like a light bulb that is always on, glowing brightly and using energy even when it's just sitting there. A Spiking Neural Network is more like a Morse code operator. It stays silent (off) most of the time and only sends a tiny "spark" (a spike) when it actually has something important to say.
The Benefit: Because it only "sparks" when necessary, it uses a fraction of the energy. The paper claims this makes it perfect for small, battery-powered medical devices (like those used in remote clinics or portable scanners).

2. The Three-View Team (The Detective Squad)

To find the tumor, the team didn't just look at the brain from one angle. They trained three separate "detectives," each looking at the brain from a different side:

Sagittal: Looking from the side.
Coronal: Looking from the front.
Axial: Looking from the top (like slicing a loaf of bread).
The Analogy: Imagine trying to guess the shape of a hidden object in a dark room. If you only have one flashlight, you might miss details. But if you have three friends shining flashlights from the left, right, and top, you get a much clearer picture.
The Result: By combining the opinions of these three "detectives," the system doesn't just find the tumor; it also knows how confident it is about its finding. If all three detectives agree, the system is very sure. If they disagree, the system knows to be cautious. This "uncertainty" is crucial for doctors to trust the AI.

3. The "Forward-Only" Training (The Efficient Student)

Training these special networks is usually very hard and expensive because it requires the computer to "rewind" and check its mistakes over and over again (like a student re-reading a whole book to find one typo).

The Innovation: The authors used a method called FPTT (Forward Propagation Through Time).
The Analogy: Instead of rewinding the tape, this method is like a student who learns as they read forward. They take a step, learn from it immediately, and move to the next step without ever looking back.
The Benefit: This makes the training process incredibly fast and saves a massive amount of computing power. The paper states this cuts the computational work (FLOPs) by 87%.

4. The Results: Fast, Cheap, and Trustworthy

The team tested their "Electric Scooter" system on real brain scan data from two major medical challenges (BraTS 2017 and 2023).

Accuracy: It found the tumors just as well as the heavy, power-hungry computers.
Reliability: Because it used the "Three-View Team," its confidence scores were very accurate. It knew when it was right and when it was unsure.
Efficiency: It used 87% less computing power than the standard methods.

Summary

In short, this paper presents a new way to map brain tumors that is lighter on energy (like a spiking neuron vs. a lightbulb), smarter at teamwork (using three views to build confidence), and cheaper to train (learning forward without rewinding). The authors argue this makes it a reliable tool for bringing high-quality medical imaging to places where big, powerful computers can't go.

Technical Summary: Reliable Brain Tumor Segmentation Based on Spiking Neural Networks with Efficient Training

Problem Statement
Accurate and reliable brain tumor segmentation is critical for clinical diagnosis and treatment planning. While Deep Neural Networks (DNNs) have achieved high performance in 3D brain tumor segmentation, their substantial computational and energy demands limit their deployment in medical Internet of Things (IoT) and Point-of-Care (PoC) systems, which often operate under strict power constraints. Furthermore, clinical integration requires models that not only delineate tumors precisely but also quantify predictive confidence (uncertainty) to ensure safety and interpretability. Existing Spiking Neural Network (SNN) approaches for uncertainty estimation often rely on static inputs repeated across time steps, a setting incompatible with the sequential, slice-wise inputs required for 3D medical image segmentation. Additionally, training SNNs for such complex tasks using standard Backpropagation Through Time (BPTT) incurs prohibitive computational and memory costs.

Methodology
The authors propose a framework centered on Spiking U-Seg-Net, an energy-efficient SNN architecture adapted for 3D brain tumor segmentation.

Network Architecture: The model follows a U-Net encoder-decoder structure but utilizes a single convolution per block to reduce parameters. It processes 3D MRI volumes (T1, T2, T1-Gd, FLAIR) by treating 2D slices from a specific anatomical view as sequential inputs over time steps. The network employs Parametric Leaky Integrate-and-Fire (PLIF) neurons with learnable membrane time constants and a surrogate gradient function ( $\sigma(x) = \frac{1}{\pi} \arctan(\pi x) + \frac{1}{2}$ ) to enable gradient-based training. The output generates three feature maps corresponding to Enhancing Tumor (ET), Tumor Core (TC), and Whole Tumor (WT).
Efficient Training (FPTT): To address the memory and computational bottlenecks of BPTT, the authors employ Forward Propagation Through Time (FPTT). This algorithm updates network parameters at each time step using an instantaneous risk function that includes a regularization term based on a running average of past weights. This approach avoids explicit backtracking through all past states, enabling scalable optimization of temporally dynamic SNNs with reduced computational cost.
Multi-View Ensemble: To enhance robustness and uncertainty estimation, the authors propose a multi-view ensemble strategy. Three independent Spiking U-Seg-Net models are trained, each processing slices from a distinct anatomical plane: sagittal, coronal, and axial. During inference, the probability maps from these three models are spatially aligned and averaged voxel-wise. This ensemble approach leverages complementary information across views to provide well-calibrated uncertainty estimates, addressing the limitations of single-view SNN models.
Loss Function: Training utilizes a hybrid loss combining Binary Cross-Entropy ( $L_{BCE}$ ) and Dice Loss ( $L_{Dice}$ ) to handle class imbalance and improve sensitivity to small tumor regions.

Key Contributions

Energy-Efficient Architecture: The introduction of Spiking U-Seg-Net, an SNN-based UNet trained with FPTT, specifically designed for brain tumor segmentation.
Robust Uncertainty Estimation: A multi-view ensemble mechanism that aggregates predictions from sagittal, coronal, and axial views, significantly improving model reliability and uncertainty calibration compared to single-view approaches.
Comprehensive Evaluation: A rigorous assessment on the BraTS 2017 and BraTS 2023 datasets demonstrating that the proposed method achieves competitive accuracy while drastically reducing computational load.

Experimental Results
The framework was evaluated on the BraTS 2017 (210 subjects) and BraTS 2023 (1,251 subjects) datasets using 5-fold cross-validation.

Segmentation Accuracy: The ensemble Spiking U-Seg-Net achieved competitive Dice ratios. On BraTS 2017, it approached the performance of DNN baselines (U-Seg-Net and U-Seg-Net+CLSTM). On BraTS 2023, it demonstrated superior performance in segmenting the Whole Tumor (WT) region compared to the baselines.
Uncertainty Calibration: The multi-view ensemble significantly improved uncertainty estimation. On BraTS 2017, the ensemble reduced the Negative Log-Likelihood (NLL) by approximately 43.8% compared to the best single-view model. On BraTS 2023, the improvement was 48.3%, indicating that the ensemble provides more reliable and well-calibrated predictive probabilities.
Computational Efficiency: The use of sparse SNN computation resulted in an 87% reduction in FLOPs (Floating Point Operations) compared to the DNN baseline when processing a single sample, highlighting the potential for low-power deployment.

Significance and Claims
The paper claims that the proposed framework demonstrates the potential of Spiking Neural Networks for reliable, low-power medical IoT and Point-of-Care systems. By integrating FPTT for efficient training and a multi-view ensemble for robust uncertainty estimation, the authors show that SNNs can deliver accurate 3D brain tumor segmentation without the high energy costs associated with traditional DNNs. The work positions SNNs as a viable foundation for uncertainty-aware medical imaging, capable of operating under strict resource constraints while maintaining clinical-grade reliability.