Abstract
Wireless Capsule Endoscopy (WCE) provides a non-invasive method for gastrointestinal imaging but is constrained by significant data transmission and battery life limitations. This paper presents a hardware-efficient segmentation method designed for on-device implementation. By utilizing a Multi-Layer Perceptron (MLP) with ternary quantized weights (-1, 0, +1) and optimized color channel selection, the proposed system eliminates the need for power-hungry digital multipliers. Experimental results demonstrate that the quantized model achieves a DICE score of 0.8403, comparable to full-precision models, enabling real-time, low-power video summarization directly within the capsule.
Introduction
Wireless Capsule Endoscopy (WCE) has revolutionized the diagnosis of gastrointestinal diseases. However, a typical procedure generates thousands of images, the majority of which depict normal tissue. Transmitting this entire high-resolution video stream to an external receiver is the primary driver of battery consumption and memory usage in WCE devices.
To address this bottleneck, researchers are exploring "in-capsule" intelligence to summarize video data before transmission. The goal is to identify abnormalities, such as bleeding, on the fly to prioritize relevant frames. However, the extremely limited hardware resources and power budget of WCE chips make standard deep-learning architectures impractical. This study proposes a streamlined, multiplication-free segmentation pipeline tailored for low-power hardware implementation.
Proposed Methodology
The authors developed a pipeline focused on reducing computational complexity without sacrificing diagnostic accuracy. The process involves three main stages: optimal color selection, patch-based feature extraction, and classification via a quantized neural network.
Optimal Color Selection
Standard RGB images contain redundant information. To minimize the input data for the neural network, the researchers evaluated various color spaces to identify which channels most effectively distinguish blood from normal tissue. They determined that a combination of Saturation (HSI), the "a" channel (CIELab), and Grayscale provides the most distinctive features for bleeding detection.
Quantized Multi-Layer Perceptron (MLP)
The core of the system is an MLP with a 40-20-8 architecture. To meet hardware constraints, the authors employed ternary quantization, restricting network weights to only three values: -1, 0, or +1. In digital logic, this transformation is significant because it replaces complex multipliers with simple sign-change operations or zero-skipping logic, drastically reducing the circuit area and power consumption required for inference.
While ReLU is a standard activation function in modern deep learning, the authors found it performed poorly in this specific quantized context. Instead, they utilized a sigmoid activation function to maintain segmentation quality.
Experimental Results
Dataset and Validation
The model was validated using a public WCE bleeding image dataset consisting of 50 images. To ensure robust results, a 5-fold cross-validation method was used. Given that bleeding pixels represent only about 0.2% of a typical WCE image, the training data was balanced to include an equal number of bleeding and non-bleeding pixels, preventing model bias toward the majority class.
Performance Metrics
The performance was evaluated using the DICE score, which measures the overlap between the predicted segmentation and the physician-provided ground truth.
- Full-Precision MLP: 0.8553 DICE score
- Quantized MLP: 0.8403 DICE score
The negligible drop in performance (approximately 1.5%) demonstrates that ternary quantization is a viable strategy for medical image segmentation in resource-constrained environments. The quantized model performed comparably to more complex Support Vector Machine (SVM) approaches while requiring a fraction of the computational overhead.
Conclusion and Future Work
This research demonstrates that a multiplication-free, quantized neural network can effectively segment bleeding regions in WCE images. By combining feature engineering (optimal color selection) with extreme weight quantization, the authors provide a blueprint for "always-on" medical AI that can operate within the strict power limits of an endoscopic capsule.
Future work will focus on expanding the model to detect other pathologies, such as polyps and ulcers, and implementing the design on an actual FPGA or ASIC to measure physical power savings and gate counts in a real-world hardware environment.
Summary updated