Core Concepts
A parallel spatial-spectral classification approach is evaluated on different high-performance computing platforms to enable real-time hyperspectral imaging for medical diagnosis.
Abstract
The paper presents a comparison of the performance and power consumption of parallel implementations of a spatial-spectral (SS) classifier for hyperspectral imaging (HSI) on various high-performance computing (HPC) platforms. The SS classifier combines spectral and spatial information to improve the classification of HSI data, which is crucial for medical applications like brain tumor identification and skin cancer detection.
The key highlights are:
The SS classifier consists of three main algorithms: Principal Component Analysis (PCA) for dimensionality reduction, Support Vector Machine (SVM) for pixel-wise classification, and K-Nearest Neighbors (KNN) for spatial filtering.
The parallel implementations are evaluated on three high-power desktop GPUs, a low-power GPU (Jetson TX2), and a low-power manycore platform (MPPA-256-N). The goal is to find a balance between processing time and energy efficiency.
For the neurosurgical use case, where real-time performance is crucial, the high-power GPUs provide the best processing times. However, for the dermatological use case where energy efficiency is more important, the low-power platforms show promising results.
The manycore platform achieves good energy efficiency by optimizing data transfers and memory usage, while the low-power GPU leverages its shared CPU-GPU memory to avoid costly data transfers.
The analysis provides insights on the trade-offs between performance and power consumption for different HPC platforms when deploying the spatial-spectral hyperspectral image classifier in medical applications.
Stats
The paper does not provide specific numerical data or metrics in the main text. However, it mentions that the HS images used in the experiments have the following characteristics:
Brain cancer images: 128 spectral bands, covering the range between 450 and 900 nm.
Dermatological images: 100 spectral bands after pre-processing.
Quotes
The paper does not contain any direct quotes that are relevant to the key points.