Tongue Tumor Detection in Medical Hyperspectral Images

Zhi Liu, Hongjun Wang, Qingli LiSensors 2011 · DOI

TL;DR

This paper introduces a novel hyperspectral imaging (HSI) system for non-invasively detecting tongue tumors. By combining an acousto-optic tunable filter (AOTF) based HSI system with a sparse representation (SR) classification algorithm, they achieve a high recognition rate of 96.5%, offering a promising tool for early cancer diagnosis.

What Problem Does This Paper Address?

Early detection of tongue cancer is crucial for improving patient outcomes, as treatment is simpler and more effective at earlier stages. Current diagnostic methods like histopathology are invasive, expensive, time-consuming, and only sample small areas of tissue. While computer vision offers non-invasive alternatives, conventional cameras capture limited spectral information. Hyperspectral imaging (HSI) offers a solution by capturing detailed spectral signatures across many narrow bands, enabling the differentiation of tissues based on their unique spectral characteristics. However, existing HSI systems for medical applications face challenges such as high temporal cost, poor spatial resolution, and difficulties in accurately delineating tumors due to tongue movement and saliva. This paper aims to overcome these limitations by developing a faster, more accurate, and non-invasive HSI system for tongue tumor detection.

Methods

The core of this research lies in the development of a Medical Hyperspectral Imaging (MHSI) system and a sophisticated classification algorithm.

  • Hyperspectral Imaging System: The system utilizes an acousto-optic tunable filter (AOTF) for rapid wavelength scanning. This solid-state technology offers advantages over traditional filter wheels or gratings, including no moving parts, high speed (up to 100 kHz), and a broad tuning range (600-1000 nm in this case). The system comprises a spectral illuminator (500W halogen lamps), an AOTF unit (Brimrose VA210), and a high-resolution camera (JAI BM141-GE) capable of capturing 81 mono-channel images with 5 nm spectral resolution. This setup allows for real-time acquisition of hyperspectral "cubes" where each pixel contains a detailed spectral signature.
  • Preprocessing: Before classification, hyperspectral images undergo denoising (using a median filter to address noise from saliva and tongue movement) and normalization. Normalization is critical to account for variations in illumination and dark current. A standard white board is used as a reference for white balancing, and dark current is captured with the shutter closed. The relative reflectance is calculated using a specific formula: $R(\lambda) = \frac{I_{raw}(\lambda) - I_{dark}(\lambda)}{I_{white}(\lambda) - I_{dark}(\lambda)}$.
  • Sparse Representation (SR) Classification: The classification algorithm is based on sparse representation. The core idea is that a hyperspectral pixel (represented as a high-dimensional vector) can be sparsely represented as a linear combination of a dictionary of training samples.
  • Dictionary Learning: A dictionary is constructed from training hyperspectral images of both noncancerous and cancerous tongue tissues. This dictionary is a matrix where columns represent spectral signatures of individual pixels.
  • Sparse Coding: For a given test pixel, the algorithm attempts to find the sparsest set of coefficients that best reconstructs the pixel's spectrum using the combined dictionary. This is formulated as an optimization problem: $\hat{\alpha} = \arg \min ||\alpha||_1 \text{ subject to } ||y - A\alpha||_2 \le \epsilon$.
  • Classification: After obtaining the sparse coefficients, the residuals are calculated for both the noncancerous and cancerous dictionaries. A decision metric $D(x) = \lg \frac{||x - A_{nc}\hat{\alpha}||_2}{||x - A_c\hat{\beta}||_2}$ is used. If $D(x) > 0$, the pixel is classified as cancerous; otherwise, it's classified as noncancerous.

Datasets

The paper introduces a custom-built tongue tumor image database. This database consists of:

  • 65 tumors.
  • 34 full tumor/partial resection/tumor bed sets used for analysis.

For evaluation, approximately 10% of the labeled samples were used for training, and 90% for testing. The specific breakdown of training and testing samples for noncancerous and cancerous classes is provided in Table 1:

Class Train Test
Noncancerous 954 8,609
Cancerous 796 7,237

Key Insights and Ideas

  • Spectral Signatures as Biomarkers: The fundamental insight is that cancerous and normal tongue tissues exhibit distinct spectral signatures in the visible and near-infrared range. These differences are likely related to underlying biochemical changes, such as protein content, as suggested by prior work.
  • AOTF for Speed and Versatility: The use of an AOTF is a key enabler, providing fast, electronically tunable spectral filtering without mechanical parts. This is crucial for real-time medical imaging and overcoming the challenges of patient movement.
  • Sparse Representation for Robust Classification: SR proves to be a powerful technique for handling the high dimensionality of hyperspectral data. Its ability to represent signals as a sparse linear combination of dictionary elements allows for effective discrimination between tissue types, even in the presence of noise and variations.
  • Dynamic Dictionary Search: Unlike methods like SVM or RVM that use fixed feature spaces, the SR-based method dynamically searches for relevant "atoms" (spectral signatures) in the training dictionary for each test pixel. This dynamic nature contributes to its robustness.
  • Non-invasive Diagnosis Potential: The system's success in achieving high accuracy without requiring tissue removal highlights its potential as a non-invasive diagnostic tool, which can significantly improve patient comfort and reduce healthcare costs.

Key Results

  • High Recognition Rate: The proposed MHSI system combined with the SR classification algorithm achieved a best recognition rate of 96.5%.
  • Effective Tumor Delineation: Figure 7 demonstrates that the system can accurately identify cancerous regions, comparable to expert labeling.
  • Performance Comparison: The SR method outperformed Support Vector Machine (SVM) and Relevance Vector Machine (RVM) in terms of overall accuracy.
  • Speed Advantage: Despite being computationally intensive, the SR algorithm achieved faster classification times (3.4 seconds) compared to SVM (7.8 seconds) and RVM (6.9 seconds) on the tested hardware.
  • Robustness: The method worked well even for tumors up to 3 mm deep and those covered with mucosa.
  • Low Error Rates: The system achieved a False Positive Rate (FPR) of 6.3% and a False Negative Rate (FNR) of 8.7%, which were lower than SVM and RVM.

Limitations and Future Work

  • Database Size: While the constructed database is valuable, the authors acknowledge the need for larger, publicly available medical HSI databases for broader validation.
  • Computational Intensity: Although faster than SVM/RVM in this specific implementation, SR algorithms can be computationally intensive, especially for very large dictionaries or complex scenes.
  • Noise and Artifacts: While preprocessing steps like median filtering and normalization are employed, challenges like glare and artifacts from saliva can still impact detection. Mathematical morphology was used as a post-processing step to fill holes caused by glare.
  • Future Work: The authors plan follow-up studies to:
  • Quantitatively grade tumors automatically based on clinicopathologic features.
  • Further investigate the spectrochemical properties of tumor tissues.
  • Explore the system's application in assessing tissue response to therapy.
  • Develop the system into a computer-aided medical diagnostic tool.

Summary updated

Abstract

A hyperspectral imaging system to measure and analyze the reflectance spectra of the human tongue with high spatial resolution is proposed for tongue tumor detection. To achieve fast and accurate performance for detecting tongue tumors, reflectance data were collected using spectral acousto-optic tunable filters and a spectral adapter, and sparse representation was used for the data analysis algorithm. Based on the tumor image database, a recognition rate of 96.5% was achieved. The experimental results show that hyperspectral imaging for tongue tumor diagnosis, together with the spectroscopic classification method provide a new approach for the noninvasive computer-aided diagnosis of tongue tumors.
BibTeX
@inproceedings{liu2011tongue,
  author = {Zhi Liu and Hongjun Wang and Qingli Li},
  title = {{Tongue Tumor Detection in Medical Hyperspectral Images}},
  booktitle = {Sensors},
  year = {2011},
  doi = {10.3390/s120100162},
}
Figure