Blog

AI-based detection of Oral Cavity Squamous Cell Carcinoma from photographic images

A Deep Learning Algorithm for Detection of Oral Cavity Squamous Cell Carcinoma from Photographic Images: A Retrospective Study

Authors: Qiuyun Fu, Yehansen Chen, Zhihang Li, Qianyan Jing, Chuanyu Hu, Han Liu, Jiahao Bao, Yuming Hong, Ting Shi, Kaixiong Li, Haixiao Zou, Yong Song, Hengkun Wang, Xiqian Wang, Yufan Wang, Jianying Liu, Hui Liu, Sulin Chen, Ruibin Chen, Man Zhang, Jingjing Zhao, Junbo Xiang, Bing Liu, Jun Jia, Hanjiang Wu, Yifang Zhao, Lin Wan, Xuepeng Xiong

Summary
This multicenter Chinese study developed a deep learning algorithm for automatic detection of oral cavity squamous cell carcinoma (OCSCC) from photographic images, aiming to create a non-invasive, rapid, and low-cost screening tool for early diagnosis.

The dataset included 7.244 clinical oral photographs collected between 2006 and 2019 from 11 hospitals. All biopsy-proven OCSCC and normal controls were included after strict quality control to remove intra-operative, postoperative, and blurred i... The algorithm, based on cascaded convolutional neural networks (CNNs), was trained on 5,775 images (development), internally validated on 401, externally on 402, and clinically tested on 666 images.

Internal validation: AUC = 0.983 | Accuracy = 91.5 % | Sensitivity = 94.9 % | Specificity = 88.7 %
Early-stage lesions (< 2 cm): AUC = 0.995 | Accuracy = 95.3 %
External validation: AUC = 0.935 | Accuracy = 84.1 %
Clinical dataset: AUC = 0.970 | Accuracy = 92.3 %

The algorithm achieved performance equivalent to that of seven oral-cancer specialists and markedly superior to medical and non-medical students.
A smartphone application (OCSCC-Detector) was developed for real-time detection and lesion scoring for clinical and remote use.

DOI: https://doi.org/10.1016/j.eclinm.2020.100558

Key Words

Oral Cancer, Squamous Cell Carcinoma, Photographic Images, Deep Learning, CNN, DenseNet121, SSD Detector, Early Detection, Smartphone App, Computer Vision, Artificial Intelligence in Dentistry

Extracted Data

Year: 2020

Modality: Clinical Photographs (Intraoral – Oral Mucosa)

Dataset: 44,409 images (11 hospitals)

Dataset Split: Development = 5,775 | Internal Validation = 401 | External = 402 | Clinical = 666
##Network Architecture: Cascaded CNN – Single Shot MultiBox Detector (SSD) + DenseNet121 classifier | transfer learning from ImageNet
##Metrics: AUC = 0.983 (internal) | 0.995 (early) | 0.935 (external) | 0.970 (clinical); Accuracy = 91.5–95.3 %

AP – Strategy: Retrospective multicenter dataset; biopsy-proven diagnoses; quality control and data augmentation (rotate, flip, brightness, scaling)

AP – Professional Qty: 21 human readers (7 specialists, 7 medical students, 7 non-medical students for comparison)

AP – Supervisor Presence: Yes (Wuhan University & collaborating centers)

AP – Experience Level: No information

AP – Expertise Area: Oral & Maxillofacial Surgery, Medical oncologists, non-medical students

AP – Tool or System: No information

Task: Classification (OCSCC vs Normal Mucosa)

Project Objective: To develop a deep learning algorithm that detects oral cancer from ordinary photographs with accuracy comparable to specialists and suitable for clinical or mobile use

Clinical Relevance

Clinical importance: Early identification of oral squamous cell carcinoma is crucial for improving survival and reducing treatment burden.

Innovation: The AI system detects OCSCC with expert-level precision from routine photographs, including early lesions < 2 cm.

Practical impact: This technology enables non-invasive, accessible, and cost-effective screening in low-resource or community settings, supporting specialists and expanding preventive care coverage.

YOUTYBE: https://youtu.be/4ozsrb2mulQ