Hyperspectral Medical Image Alignment Method for Extracting Spectral Features

Abstract

Hyperspectral imaging is a powerful tool in dermatology. It collects skin data across many wavelengths, from ultraviolet to infrared, creating a 3D data cube. This allows researchers to extract specific spectral features of skin lesions for further analysis.

However, in vivo medical imaging faces a major practical problem: patient movement and hand tremor during the scanning process. Because the camera takes 151 consecutive frames over time, the skin lesion shifts its position from frame to frame. If we extract spectral data from a single pixel coordinate without correction, we will capture different parts of the tissue at different wavelengths, making the extracted features invalid.

This paper presents an automated pipeline for the spatial alignment of hyperspectral skin image cubes. We implemented a hybrid tracking method that combines image processing and physics-based prediction. First, we use Difference of Gaussians (DoG) filtering to highlight skin texture, reducing the impact of changing brightness across the spectrum. Then, we apply normalized cross-correlation (NCC) combined with a Kalman filter. The Kalman filter predicts the movement of the lesion when the visual contrast drops significantly, especially in the infrared range. Finally, we use affine transformations to shift all frames to match a single reference frame.

The proposed method was evaluated on a dataset of hyperspectral images containing various skin conditions. Experimental results, confirmed by NCC and Structural Similarity (SSIM) metrics, show a significant improvement in spatial consistency across all wavelengths. This alignment ensures that the extracted spectral features belong to the exact same tissue area, providing a reliable foundation for future classification.

Speaker

Artem Nepovinnov
Samara National Research University named after Academician S.P. Korolev
Russian Federation

Discussion

Ask question