Quantitative Detection Of Head And Neck Cancer Using Hyperspectral Imaging

Head and neck cancer(HNC)is the sixth most common cancer worldwide.The early diagnosis of HNC can avoid high stage of the disease,and the accurate margin detection during the tumor resection is vital.Hyperspectral imaging(HSI)is a noncontact,non-invasive,non-ionizing,and label-free optical imaging technology,and it is emerging as one of the research hotspots in medical imaging.However,the spectral differences between cancerous and normal tissues are subtle,therefore it is hard to achieve accurate diagnosis using original spectra of tissue.On the other hand,it is necessary for HSI to fulfill fast quantitative cancer detection in order to be utilized as intraoperative image guidance and replace the time-consuming intraoperative pathologist consultation procedure.This thesis studies and develops quantitative HNC detection techniques using HSI in aspects of in vivo tumor detection,ex vivo tumor margin assessment,and cancer detection in histopathological slides.The main contents are summarized as below.(1)A hyperspectral image feature extraction method based on one-dimensional discrete wavelet transform(DWT)was proposed.A five-order DWT was applied to the reflectance spectra of in vivo tumor and normal tissues of mice,and all outputted wavelet coefficients were concatenated in sequence to form a spectral feature of each spectrum.Machine learning method was employed to classify the reflectance spectra and wavelet-based features,and the classification results using features generated with different mother wavelets were compared.The best mother wavelet option for the in vivo tumor detection application was obtained.The classification got 87.5% accuracy,90.0% specificity,and 83.1% sensitivity.The proposed feature extraction method outperformed original reflectance spectra for quantitative in vivo tumor analysis and improves the detection efficiency.(2)A tumor margin assessment method based on semantic segmentation network was proposed.Based on the characteristics of the hyperspectral images of ex vivo HNC tissue,a U-Net was optimized by modifying the architecture,as well as applying batch normalization,regularization,weight initialization,etc.The margin detection error was calculated using 95% Hausdorff distance(95% HD).The proposed method could achieve pixel-level classification in hyperspectral images of ex vivo surgical specimens,and got 0.90 AUC,0.82 accuracy,0.82 sensitivity,0.74 specificity,as well as 1.8 mm average 95% HD in the testing image dataset.The classification of each specimen took less than 20 ms.The comparison between classification experiments using HSI and RGB indicates that HSI has better ability for tumor detection.Compared to other commonly used classification networks,the proposed method can achieve faster image classification with higher precision,and no image reconstruction is needed after classification.Therefore,it can serve as an intraoperative tool to reduce the surgical time and to improve the surgical outcome.(3)A whole-slide cancer detection method based on hyperspectral microscopic imaging was developed.Firstly,a semi-automatic whole-slide scanning hyperspectral microscopic imaging system was developed,with which a hyperspectral whole-slide histology image dataset was obtained.Secondly,an unsupervised hyperspectral image super-resolution reconstruction model was proposed,which fuses low-resolution hyperspectral images and high-resolution RGB histology images to generate high spatial-and spectral-resolution hyperspectral images.It not only achieves high-quality reconstruction in different scales,but also improves the quality of generated hyperspectral images.Thirdly,a deep CNN model based on Inception modules was established for whole-slide image classification and train with RGB images,original and generated hyperspectral images,respectively.Results obtained using either type of hyperspectral data were better than the RGB images,while the reconstructed hyperspectral data using the proposed super-resolution network even outperformed the original hyperspectral data.The proposed method can significantly reduce the cost and acquisition,improve the stability and reproducibility of histopathological diagnosis,and reduces the diagnostic time.(4)A cancer detection method in histological slides based on nucleus classification was developed.Firstly,a dual-modality microscopic imaging system was built and calibrated,with which both hyperspectral images and RGB histology images can be acquired simultaneously.Then,a semi-automatic nucleus segmentation method was proposed,which fulfilled fast nucleus image segmentation by applying principal component analysis to hyperspectral histological images and a hard threshold to the difference between the first and second principal components.Next,a feedforward neural network was developed for the classification of nucleus patches,which obtained 0.89 validation accuracy and 0.82 testing accuracy.Finally,imagewise cancer region identification was fulfilled according to the detection ratio of cancerous nuclei in the images with a 0.95 overall accuracy.The comparison among the classification results using the average transmittance spectra,RGB patches,and HSI patches of nuclei indicates that the spatial and spectral information in HSI can improve the detection of cancerous nuclei.The proposed simple-yet-effective method can improve the cancer detection efficiency in suspicious regions,especially beneficial to facilitate automatic head and neck cancer detection when combined with the abovementioned whole-slide cancer detection method.