Automatic detection of head and neck squamous cell carcinoma on histologic slides using hyperspectral microscopic imaging.

Significance: Automatic, fast, and accurate identification of cancer on histologic slides has many applications in oncologic pathology.

Aim: The purpose of this study is to investigate hyperspectral imaging (HSI) for automatic detection of head and neck cancer nuclei in histologic slides, as well as cancer region identification based on nuclei detection.

Approach: A customized hyperspectral microscopic imaging system was developed and used to scan histologic slides from 20 patients with squamous cell carcinoma (SCC). Hyperspectral images and red, green, and blue (RGB) images of the histologic slides with the same field of view were obtained and registered. A principal component analysis-based nuclei segmentation method was developed to extract nuclei patches from the hyperspectral images and the coregistered RGB images. Spectra-based support vector machine and patch-based convolutional neural networks (CNNs) were implemented for nuclei classification. The CNNs were trained with RGB patches (RGB-CNN) and hyperspectral patches (HSI-CNN) of the segmented nuclei and the utility of the extra spectral information provided by HSI was evaluated. Furthermore, cancer region identification was implemented by image-wise classification based on the percentage of cancerous nuclei detected in each image.

Results: RGB-CNN, which mainly used the spatial information of nuclei, resulted in a 0.81 validation accuracy and 0.74 testing accuracy. HSI-CNN, which utilized the spatial and spectral features of the nuclei, showed significant improvement in classification performance and achieved 0.89 validation accuracy as well as 0.82 testing accuracy. Furthermore, the image-wise cancer region identification based on nuclei detection could generally improve the cancer detection rate.

Conclusions: We demonstrated that the morphological and spectral information contribute to SCC nuclei differentiation and that the spectral information within hyperspectral images could improve classification performance.