A domain knowledge enhanced yield based deep learning classifier identifies perineural invasion in oral cavity squamous cell carcinoma

Li-Yu Lee; Cheng-Han Yang; Yu-Chieh Lin; Yu-Han Hsieh; Yung-An Chen; Margaret Dah-Tsyr Chang; Yen-Yin Lin; Chun-Ta Liao

doi:10.3389/fonc.2022.951560

A domain knowledge enhanced yield based deep learning classifier identifies perineural invasion in oral cavity squamous cell carcinoma

Front Oncol. 2022 Oct 24:12:951560. doi: 10.3389/fonc.2022.951560. eCollection 2022.

Authors

Li-Yu Lee¹, Cheng-Han Yang¹, Yu-Chieh Lin^{2

3}, Yu-Han Hsieh², Yung-An Chen², Margaret Dah-Tsyr Chang², Yen-Yin Lin², Chun-Ta Liao⁴

Affiliations

¹ Department of Pathology, Chang Gung Memorial Hospital and Chang Gung University, Taoyuan, Taiwan.
² Department of Strategic Technology, JelloX Biotech Inc., Hsinchu, Taiwan.
³ Department of Power Mechanical Engineering, National Tsing Hua University, Hsinchu, Taiwan.
⁴ Department of Otorhinolaryngology, Head and Neck Surgery, Chang Gung Memorial Hospital and Chang Gung University, Taoyuan, Taiwan.

Abstract

Background: Perineural invasion (PNI), a form of local invasion defined as the ability of cancer cells to invade in, around, and through nerves, has a negative prognostic impact in oral cavity squamous cell carcinoma (OCSCC). Unfortunately, the diagnosis of PNI suffers from a significant degree of intra- and interobserver variability. The aim of this pilot study was to develop a deep learning-based human-enhanced tool, termed domain knowledge enhanced yield (Domain-KEY) algorithm, for identifying PNI in digital slides.

Methods: Hematoxylin and eosin (H&E)-stained whole-slide images (WSIs, n = 85) were obtained from 80 patients with OCSCC. The model structure consisted of two parts to simulate human decision-making skills in diagnostic pathology. To this aim, two semantic segmentation models were constructed (i.e., identification of nerve fibers followed by the diagnosis of PNI). The inferred results were subsequently subjected to post-processing of generated decision rules for diagnostic labeling. Ten H&E-stained WSIs not previously used in the study were read and labeled by the Domain-KEY algorithm. Thereafter, labeling correctness was visually inspected by two independent pathologists.

Results: The Domain-KEY algorithm was found to outperform the ResnetV2_50 classifier for the detection of PNI (diagnostic accuracy: 89.01% and 61.94%, respectively). On analyzing WSIs, the algorithm achieved a mean diagnostic accuracy as high as 97.50% versus traditional pathology. The observed accuracy in a validation dataset of 25 WSIs obtained from seven patients with oropharyngeal (cancer of the tongue base, n = 1; tonsil cancer, n = 1; soft palate cancer, n = 1) and hypopharyngeal (cancer of posterior wall, n = 2; pyriform sinus cancer, n = 2) malignancies was 96%. Notably, the algorithm was successfully applied in the analysis of WSIs to shorten the time required to reach a diagnosis. The addition of the hybrid intelligence model decreased the mean time required to reach a diagnosis by 15.0% and 23.7% for the first and second pathologists, respectively. On analyzing digital slides, the tool was effective in supporting human diagnostic thinking.

Conclusions: The Domain-KEY algorithm successfully mimicked human decision-making skills and supported expert pathologists in the routine diagnosis of PNI.

Keywords: artificial intelligence; deep learning; digital pathology; oral cavity squamous cell carcinoma; perineural invasion.