AUTHOR=Xu Ziang , Peng Jiakuan , Zeng Xin , Xu Hao , Chen Qianming 

TITLE=High-Accuracy Oral Squamous Cell Carcinoma Auxiliary Diagnosis System Based on EfficientNet

JOURNAL=Frontiers in Oncology

VOLUME=Volume 12 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2022.894978

DOI=10.3389/fonc.2022.894978

ISSN=2234-943X

ABSTRACT=It is important to diagnose the grade of squamous cell carcinoma (OSCC), while the current evaluation of the biopsy slide still mainly depends on the manual operation of the pathologists. The workload of manual evaluation is large, and the results are greatly affected by the subjectivity of the pathologists. In recent years, with the development and application of deep learning, automatic evaluation of biopsy slide is gradually being applied on medical diagnosis, and it has shown good results. Therefore, a new OSCC auxiliary diagnostic system was proposed to automatically and accurately evaluate the patients’ tissue slide. This is the first study which compared the effects of different resolutions on the results. The OSCC tissue slides from The Cancer Genome Atlas (TCGA, n=697) and our independent datasets(n=337) were used for model training and verification. The test results on tiles show that the accuracy of 93.1%for 20x resolution (n=306,134) is higher than 90.9% for 10x (n= 154,148) and 89.3% for 40x (n= 890,681). The accuracy of the new system based on EfficientNet, which used to evaluate the tumor grade of biopsy slide, reached 98.1% (95% confidence interval [CI]: 97.1% to 99.1%) and area under the receiver operating characteristic curve (AUROC) reached 0.998 (95%CI: 0.995 to 1.000) in TCGA dataset. When verifying the model in independent image dataset, the accuracy still reached 91.4% (95% CI: 88.4% to 94.4%, at 20x) and the AUROC reached 0.992 (95%CI: 0.982 to 1.000). It may benefit the oral pathologists in reducing certain repetitive and time-consuming tasks, improve the efficiency of diagnosis, and facilitate the further development of computational histopathology.