Development of deep learning methods for head and neck cancer detection in hyperspectral imaging and digital pathology for surgical guidance

Abstract

Surgeons performing routine cancer resections utilize palpation and visual inspection, along with time-consuming microscopic tissue analysis, to ensure removal of cancer. Despite this, inadequate surgical cancer margins are reported for up to 10-20% of head and neck squamous cell carcinoma (SCC) operations. There exists a need for surgical guidance with optical imaging to ensure complete cancer resection in the operating room. The objective of this dissertation is to evaluate hyperspectral imaging (HSI) as a non-contact, label-free optical imaging modality to provide intraoperative diagnostic information. For comparison of different optical methods, autofluorescence, RGB composite images synthesized from HSI, and two fluorescent dyes are also acquired and investigated for head and neck cancer detection. A novel and comprehensive dataset was obtained of 585 excised tissue specimens from 204 patients undergoing routine head and neck cancer surgeries. The first aim was to use SCC tissue specimens to determine the potential of HSI for surgical guidance in the challenging task of head and neck SCC detection. It is hypothesized that HSI could reduce time and provide quantitative cancer predictions. State-of-the-art deep learning algorithms were developed for SCC detection in 102 patients and compared to other optical methods. HSI detected SCC with a median AUC score of 85%, and several anatomical locations demonstrated good SCC detection, such as the larynx, oropharynx, hypopharynx, and nasal cavity. To understand the ability of HSI for SCC detection, the most important spectral features were calculated and correlated with known cancer physiology signals, notably oxygenated and deoxygenated hemoglobin. The second aim was to evaluate HSI for tumor detection in thyroid and salivary glands, and RGB images were synthesized using the spectral response curves of the human eye for comparison. Using deep learning, HSI detected thyroid tumors with 86% average AUC score, which outperformed fluorescent dyes and autofluorescence, but HSI-synthesized RGB imagery performed with 90% AUC score. The last aim was to develop deep learning algorithms for head and neck cancer detection in hundreds of digitized histology slides. Slides containing SCC or thyroid carcinoma can be distinguished from normal slides with 94% and 99% AUC scores, respectively, and SCC and thyroid carcinoma can be localized within whole-slide images with 92% and 95% AUC scores, respectively. In conclusion, the outcomes of this thesis work demonstrate that HSI and deep learning methods could aid surgeons and pathologists in detecting head and neck cancers.