Professor Yang Qintai's team of the Third Affiliated Hospital of Sun Yat-sen University made a new breakthrough in the diagnosis of nasal polyps using artificial intelligence

?

The team of Professor Yang Qintai (corresponding author) from the Department of Otorhinolaryngology-Head and Neck Surgery of the Third Affiliated Hospital of Sun Yat-sen University (SYSU), published their latest research entitled “Expert-level Diagnosis of Nasal Polyps Using Deep Learning on Whole-slide Imaging” in the?Journal of Allergy and Clinical Immunology?(IF = 14.11, the No.1 most-cited allergy/immunology journal) On December 9, 2019. This study was also contributed by Professor Han Lanqing (co-corresponding author) and Ren Yong (co-first author, Tsinghua Pearl River Delta Research Institute, Professor Chen Jianning (co-first author, Department of Pathology of the Third Affiliated Hospital of SYSU), Professor Sun Yueqi (the First Affiliated Hospital of SYSU) and Professor Hong Haiyu (the Fifth Affiliated Hospital of SYSU).?

The prognosis of different pathological types of nasal polyps is obviously different. Among them, eosinophilic nasal polyps are the most difficult to diagnose and treat, and their pathological classification is conducive to individualized precise treatment and prognosis. Existing diagnosis and classification methods of nasal polyps are based on the patient's pathological slides and 10 random high-power fields (HPFs) are selected from thousands of HPFs for evaluation and diagnosis. Sampling leads to bias in accuracy. This error can be avoided if the whole slides were counted and evaluated by the pathologist, but the workload is very large and it is difficult for the primary hospital to carry out. With the rapid development of deep learning (DL) and Whole-slide Imaging (WSI) in recent years, combination of the two technologies can perfectly solve this series of problems.

Figure: Schematic of AICEP

The main members of the team (Wu Qingwu, Deng Huiyi, Zheng Rui, Huang Xuekun and Yuan Lianxiong) retrospectively collected 1,465 nasal polyp slides in their hospital, screened out 195 for WSI, labeled ROI, yielded 26,589 high magnification patches with 1000 * 1000 resolution, and established an artificial intelligence evaluation platform (AICEP) based on DL.

Next, they used AICEP to perform internal and external verification on 12 patients of the Third Affiliated Hospital of SYSU, and 16 patients of the First and Fifth Hospitals of SYSU. The diagnostic results obtained were completely consistent with the actual results of the patients. However, the method of extracting 10 random HPFs for diagnosis may have misdiagnosed 2 and 4 patients, respectively. This showed that the excellent sensitivity and specificity of AICEP, and it had good external generalization ability.

In terms of interpretability, the team used feature visualization technology to distinguish eosinophils from non-eosinophils, and the results were consistent with pathological morphological features. At the same time, the team members also compared the time-consuming of diagnosis. It took an average of 12.7 minutes to diagnose a patient with nasal polyps using the 10 random HPFs and 148.6 minutes for a whole slide. However, AICEP needed only 5.4 minutes. In the next step, the AICEP platform will be modularized and used in the cloud platform, and the expert-level diagnosis of nasal polyp will be expanded to primary medical institutions.

This achievement was also a research result since the first "Medical Artificial Intelligence Center" of the Guangdong-Hong Kong-Macao Greater Bay Area was jointly established in April 2019 by the Third Affiliated Hospital of SYSU and the Tsinghua Pearl River Delta Research Institute. The medical artificial intelligence center is based on the abundant medical resources under the development pattern of "One Body, Two Wings, Three Cities and Four Hospitals" in the Third Affiliated Hospital of SYSU. Using the advantages of multiple discipline groups, the Tsinghua Pearl River Delta Research Institute and Tsinghua University, three major platforms were developed rapidly which include Medical Imaging Assistance Diagnosis and Treatment Platform, Artificial Intelligence Technology Service Platform, and Professional and Technical Personnel Training Platform. In all, this center is planned to be a demonstration base for medical artificial intelligence applications in the Guangdong-Hong Kong-Macao Greater Bay Area within 5 years.

Link to the paper:?https://doi.org/10.1016/j.jaci.2019.12.002