Advanced imaging and data mining technologies for medical and food safety applications

Thumbnail Image


Publication or External Link






As one of the most fast-developing research areas, biological imaging and image analysis receive more and more attentions, and have been already widely applied in many scientific fields including medical diagnosis and food safety inspection. To further investigate such a very interesting area, this research is mainly focused on advanced imaging and pattern recognition technologies in both medical and food safety applications, which include 1) noise reduction of ultra-low-dose multi-slice helical CT imaging for early lung cancer screening, and 2) automated discrimination between walnut shell and meat under hyperspectral florescence imaging.

     In the medical imaging and diagnosis area, because X-ray computed tomography (CT) has been applied to screen large populations for early lung cancer detection during the last decade, more and more attentions have been paid to studying low-dose, even ultra-low-dose X-ray CTs. However, reducing CT radiation exposure inevitably increases the noise level in the sinogram, thereby degrading the quality of reconstructed CT images. Thus, how to reduce the noise levels in the low-dose CT images becomes a meaningful topic. In this research, a nonparametric smoothing method with block based thin plate smoothing splines and the roughness penalty was introduced to restore the ultra-low-dose helical CT raw data, which was acquired under 120 kVp / 10 mAs protocol. The objective thorax image quality evaluation was first conducted to assess the image quality and noise level of proposed method. A web-based subjective evaluation system was also built for the total of 23 radiologists to compare proposed approach with traditional sinogram restoration method. Both objective and subjective evaluation studies showed the effectiveness of proposed thin-plate based nonparametric regression method in sinogram restoration of multi-slice helical ultra-low-dose CT.

     In food quality inspection area, automated discrimination between walnut shell and meat has become an imperative task in the walnut postharvest processing industry in the U.S. This research developed two hyperspectral fluorescence imaging based approaches, which were capable of differentiating walnut small shell fragments from meat. Firstly, a principal component analysis (PCA) and Gaussian mixture model (PCA-GMM)-based Bayesian classification method was introduced. PCA was used to extract features, and then the optimal number of components in PCA was selected by a cross-validation technique. The PCA-GMM-based Bayesian classifier was further applied to differentiate the walnut shell and meat according to the class-conditional probability and the prior estimated by the Gaussian mixture model. The experimental results showed the effectiveness of this PCA-GMM approach, and an overall 98.2% recognition rate was achieved. Secondly, Gaussian-kernel based Support Vector Machine (SVM) was presented for the walnut shell and meat discrimination in the hyperspectral florescence imagery. SVM was applied to seek an optimal low to high dimensional mapping such that the nonlinear separable input data in the original input data space became separable on the mapped high dimensional space, and hence fulfilled the classification between walnut shell and meat. An overall recognition rate of 98.7% was achieved by this method.

     Although the hyperspectral fluorescence imaging is capable of differentiating between walnut shell and meat, one persistent problem is how to deal with huge amount of data acquired by the hyperspectral imaging system, and hence improve the efficiency of application system. To solve this problem, an Independent Component Analysis with k-Nearest Neighbor Classifier (ICA-kNN) approach was presented in this research to reduce the data redundancy while not sacrifice the classification performance too much. An overall 90.6% detection rate was achieved given 10 optimal wavelengths, which constituted only 13% of the total acquired hyperspectral image data. In order to further evaluate the proposed method, the classification results of the ICA-kNN approach were also compared to the kNN classifier method alone. The experimental results showed that the ICA-kNN method with fewer wavelengths had the same performance as the kNN classifier alone using information from all 79 wavelengths. This demonstrated the effectiveness of the proposed ICA-kNN method for the hyperspectral band selection in the walnut shell and meat classification.