Abstract
We present a novel pixel-level spectra based multi-layer perceptron (MLP) to discriminate regions of biomedical multispectral imaging data into two categories: tissue and non-tissue. The spectra used for this study are 740nm, 780nm, 850nm, and 945nm as these wavelengths are on either side of the isosbestic point for oxyhemoglobin and deoxyhemoglobin; absorbers that are common in all healthy tissues. An MLP is trained using multispectral data from 12 human subjects and 12 non-tissue objects. The MLP is tested on three multispectral challenge image sets, from which the accuracy, sensitivity, and specificity of the model yield results of 91.3% (+/-0.2%), 98.1% (+/-0.3%), and 88.5% (+/- 0.3%) respectively.
Introduction
In image-guided surgery and automated diagnostics, a critical first step is segmentation—the process of distinguishing human tissue from background elements like surgical tools, drapes, and other equipment. While complex deep learning models can perform this task, their high computational cost often introduces latency, which is unacceptable in real-time medical applications. This paper proposes a streamlined, pixel-level approach using Near-Infrared (NIR) Multispectral Imaging (MSI) and a simple neural network to act as a high-speed filter, focusing more intensive diagnostic algorithms only on relevant tissue regions.
Methods
The proposed method combines the biological specificity of MSI with the computational efficiency of a Multi-layer Perceptron (MLP).
Wavelength Selection
Four NIR wavelengths (740nm, 780nm, 850nm, and 945nm) were chosen. These wavelengths flank the isosbestic point for oxyhemoglobin and deoxyhemoglobin. Because these absorbers are universal in healthy human tissue, they provide a robust biological signature that is independent of visual shape or texture.
MLP Architecture
The model utilizes a simple architecture designed for pixel-by-pixel processing. It consists of an input layer with 4 nodes (one for each wavelength), a hidden layer with 4 nodes using sigmoid activation, and a single output node representing the probability of the pixel being tissue (1) or non-tissue (0).
Training
The network was trained using a back-propagation algorithm. The training process aimed for a set error convergence of 1%, ensuring the model learned the spectral differences between tissue and non-tissue classes effectively.
Datasets and Evaluation
The study utilized a clinical MSI dataset for both training and validation.
- Training Data: Collected from 12 human subjects (forearm regions) and 12 diverse non-tissue objects.
- Data Augmentation: To enhance robustness, the dataset was augmented with synthetic data points generated within the standard deviation of the original sampled pixel spectra.
- Testing: Performance was evaluated using three multispectral "challenge" image sets containing complex scenes with both tissue and non-tissue objects.
Results
The MLP model demonstrated high performance across key metrics:
- Accuracy: 91.3% (± 0.2%)
- Sensitivity: 98.1% (± 0.3%)
- Specificity: 88.5% (± 0.3%)
The high sensitivity is particularly valuable in medical contexts, as it ensures that actual tissue is rarely missed. Qualitative analysis indicated that errors primarily occurred at object boundaries or in areas with significant hair (as seen in the forearm images), which can interfere with the spectral signature.
Conclusion
This research highlights the potential of using simple, pixel-level spectral analysis for real-time tissue detection. By prioritizing biological signatures over complex spatial patterns, the MLP approach offers a computationally efficient alternative to heavy deep learning models. Future work will explore incorporating spatial context through graphical models like Markov Random Fields to improve specificity and reduce errors at object edges.
Summary updated