Since 2012, the research team has begun establishing clinical data databases and tissue banks for various diseases, including pituitary adenomas, gliomas, metastatic tumors, epilepsy, cerebral hemorrhage, hydrocephalus, and others. These clinical data have laid a solid foundation for the successful implementation of this project.

In 2015, led by Professor Renzhi Wang, the academic leader of the research institute, China's first multicenter database on pituitary diseases was established under the National Population and Health Science Data Sharing Platform, based on the China Pituitary Disease Collaboration Group - the China Pituitary Disease Registry Center (cpdrn.pumch.cn) was officially put into use. As of now, it has covered more than 59 hospitals nationwide, with over 45,000 cases, making it the largest pituitary disease database in the world.

1) Research results and applications related to electronic medical record text data of pituitary adenomas: Team members have mastered the structuring and extraction methods of electronic medical record text data for pituitary adenomas and have conducted systematic application research combining machine learning methods. Currently, for directions such as treatment effects and recurrence in pituitary adenoma patients, pre-experimental studies have been conducted with related results: Predicting the long-term prognosis and potential recurrence of patients with pituitary ACTH adenomas (Cushing's disease). From the electronic medical record database of pituitary adenomas, 17 text features of 354 postoperative hormone-relieved patients with pituitary ACTH adenomas were collected. Using seven machine learning algorithms, a model to predict the recurrence of Cushing's disease patients was built, aiming to guide the postoperative follow-up plan for patients with Cushing's disease. Related results have been published in the Neuroendocrinology journal (2019;108:201-210, corresponding author, IF= 6.804). Preoperative prediction of surgical efficacy for patients with pituitary GH adenomas (acromegaly). From the electronic medical record database of pituitary adenomas, 12 preoperative text features of 668 patients with pituitary GH adenomas were collected. Using six machine learning algorithms, a model to predict the surgical efficacy for patients with acromegaly was built, aiming to guide the formulation of preoperative treatment plans and address clinical questions of whether to choose medication before surgery or direct surgery. Related results have been published in the Endocrine journal (2019 Oct 30. doi: 10.1007/s12020-019-02121-6).

2) Research results and applications related to imaging data of pituitary adenomas:

Team members have conducted systematic research on computational methods and applications based on radiomics. Currently, for directions such as preoperative diagnosis, treatment choice, and prognosis prediction in pituitary adenoma patients, pre-experimental studies have been conducted with related results as follows:

Preoperative assessment of surgical efficacy for patients with invasive functional pituitary adenomas. After collecting imaging data of 163 patients with invasive functional pituitary adenomas and extracting features, the SelectKBest-RFE algorithm was used to identify seven imaging features related to prognosis. A model to predict the prognosis for patients with invasive functional pituitary adenomas was built, aiming to guide the formulation of preoperative treatment plans. Related results have been published in the European Journal of Radiology (2019:108647). The team has mastered multiple techniques, including delineation of the region of interest for pituitary adenoma lesions, feature extraction, feature selection, model construction, and validation.

Preoperative judgment of tumor texture in patients with pituitary GH adenomas (acromegaly). After collecting imaging data of 158 patients with pituitary GH adenomas and extracting features using PyRadiomics, the elastic net algorithm was used to identify four imaging features related to the texture of acromegaly. The model's performance was successfully validated in a prospective study, and it successfully guided the formulation of preoperative treatment plans. Related results have been published in the Frontiers in Endocrinology journal (2019;10:588).

Assessing the sensitivity to radiotherapy in patients with pituitary GH adenomas (acromegaly). After collecting imaging data of 57 patients with pituitary GH adenomas and extracting features using PyRadiomics, the LOOCV leave-one-out cross-validation algorithm was used to identify six imaging features related to the sensitivity to radiotherapy in patients with acromegaly. The radiotherapy sensitivity model successfully guided the choice and formulation of radiotherapy plans for patients. Related results have been published in the Frontiers in Endocrinology journal (2019;10:403, corresponding author, IF= 3.634).

3) Facial recognition of pituitary adenomas:

Facial recognition technology is already very mature. Pituitary adenomas like Cushing's disease and acromegaly are neuroendocrine diseases with distinct facial features, caused by ACTH and GH-secreting pituitary adenomas. Over the past decade, the research team has accumulated a large amount of case data for pituitary adenomas. Therefore, using these photos, mathematical models for such diseases can be established through facial recognition and machine learning for disease screening. The dataset includes: facial frontal photos of Cushing's disease patients: 814, acromegaly patients: 1131, other types of patients: 5642, Deepglint database: 4938, FERET facial dataset: 2018. The neural network used is based on the Google Inception-v3 network that has been trained to convergence on ImageNet, with certain modifications. After validation, this neural network can diagnose Cushing's disease and acromegaly patients well.

4) DUCG auxiliary diagnosis system: The DUCG knowledge base includes 15 types of pituitary-related symptoms, signs, laboratory examinations, radiological examinations, etc., with interpretable diagnostic results. After case validation, the accuracy rate is over 95% (20 main complaint symptoms, accuracy rate 95%). The knowledge base includes: cough with phlegm, fever with rash, difficulty breathing, abdominal pain, diarrhea, hematemesis, nasal congestion, nasal bleeding, hematochezia, nausea and vomiting, joint pain, hemoptysis, fever, lower urinary tract symptoms (including hematuria, urinary frequency, urgency and pain, oliguria, anuria and polyuria, urinary incontinence, difficulty urinating). Additionally, 17 knowledge bases for main complaint symptoms have been completed and are being tested, including: chest pain, consciousness disturbance, jaundice, anemia, edema, obesity, emaciation, cyanosis, sore throat, abnormal sexual development, lymphadenopathy, palpitations, tremors, pediatric fever, gynecological-related diseases (3 main complaint symptoms), neurosurgery specialty (including saddle area diseases, etc.). Seven knowledge bases for main complaint symptoms are under construction, including: neck and back pain, dizziness, headache, constipation, rash, dysphagia, limb numbness. The system already covers common diseases in general practice, including over 2000 diseases, and continues to expand. The DUCG system is already in use in over 500 health and medical institutions. 

Based on the research team's work achievements in the diagnosis and screening of pituitary adenomas, the institute will integrate existing data platforms and systems to develop an intelligent auxiliary diagnosis platform and corresponding APP system for patient screening.