Tongue Tumor Detection in Hyperspectral Images Using Deep Learning Semantic Segmentation.

Area of Science:

• Medical Imaging
• Computational Pathology
• Surgical Technology

Background:

• Real-time tumor segmentation using hyperspectral imaging (HSI) is challenging.
• Accurate tumor detection is crucial for effective surgical intervention.

Purpose of the Study:

• To propose and evaluate deep learning semantic segmentation methods for tongue tumor detection in HSI data.
• To compare these methods with existing deep learning algorithms.
• To investigate the impact of channel selection and the use of both visual (VIS) and near-infrared (NIR) spectra.

Main Methods:

• Developed deep learning semantic segmentation models for HSI data.
• Implemented channel selection and incorporated spatial tissue context.
• Utilized both VIS and NIR spectral bands for analysis.
• Compared model performance against established deep learning algorithms.

Main Results:

• Achieved strong performance in segmenting tongue squamous cell carcinoma using HSI.
• Demonstrated the effectiveness of channel selection over using full spectra.
• Showcased the crucial role of NIR spectra in specific tumor detection cases.
• Obtained high average Dice coefficient and area under the ROC curve on clinical data.

Conclusions:

• Deep learning semantic segmentation with HSI shows significant potential for real-time surgical guidance.
• Channel selection and the inclusion of NIR spectra enhance tumor detection accuracy.
• HSI technology, augmented by deep learning, can serve as a valuable tool in surgery and digital pathology.