Purpose: Histological grading of squamous cell carcinoma (SCC) of the head and neck has been correlated with disease prognosis and patient survival. SCC is graded from well, moderately, to poorly differentiated (grades 1, 2, and 3, respectively). There is substantial inter-pathologist agreement for determining of grade, with a kappa statistic of 0.6 to 0.65. The objective of this project is to determine the ability of label-free, non-contact hyperspectral imaging (HSI) for determining histological grade of SCC in gross-level ex-vivo tissue specimens.
Methods: We collected a dataset of patients with head and neck SCC undergoing surgical cancer resection and imaged the ex-vivo tissue specimens with HSI (N=53 patients, N=67 tissue specimens). We implemented a convolutional neural network to classify the 3 SCC grades, and tested on a subset of 12 patients, 16 tissue specimens to determine performance.
Results: Overall accuracy for predicted grade of tissue specimens was 69%, and agreement of 0.5 kappa statistic (where 0 corresponds to random guess, and +1 corresponds to perfect agreement). Two out of three grade 1 specimens were predicted correctly (one incorrectly predicted as grade 2). Six out of seven grade 2 specimens were predicted correctly (one incorrectly predicted as grade 1). And, three out of six grade 3 specimens were predicted correctly (three incorrectly predicted as grade 1).
Conclusion: Our preliminary results show that label-free hyperspectral imaging may hold potential for automatic grading of head and neck SCC of gross-level tissue specimens.