Intelligent Screening and Precision Diagnosis of Prostate Cancer Based on Multimodal Data

Overview

This project aims to develop a precision screening and diagnostic solution for prostate cancer based on multimodal artificial intelligence, focusing on addressing the diagnostic challenge in patients within the PSA "gray zone" of 4-10 ng/mL. The project will integrate multidimensional information including ctDNA liquid biopsy, routine laboratory data, and prostate ultrasound images to develop three models: a ctDNA-based multimodal AI prediction model, a routine laboratory data-assisted decision model, and an ultrasound image AI-assisted diagnostic model. On this basis, a multimodal AI fusion decision system will be established to automatically generate individualized risk assessment reports and diagnostic recommendations. Additionally, a closed-loop mechanism of "clinical use - data feedback - model optimization" will be constructed to continuously iterate model parameters using pathological gold standards, thereby improving predictive accuracy in our hospital population. The project will form a generalizable precision diagnostic workflow, reduce unnecessary biopsies in "gray zone" patients, and provide an implementable in-hospital solution for precision medicine in prostate cancer.

Background: Prostate cancer (PCa) is the second most common malignancy in men worldwide. In China, the average annual growth rate of PCa incidence is as high as 7.2%. Current diagnostic pathways rely on transrectal ultrasound (TRUS)-guided prostate biopsy. However, serum PSA, the main decision-making indicator for biopsy, is not cancer-specific and has severely insufficient specificity. Many men with elevated PSA undergo unnecessary invasive biopsies. The diagnostic challenge is particularly prominent in the PSA "gray zone" of 4-10 ng/mL. Objectives: This study aims to develop a precision screening and diagnostic solution for prostate cancer based on multimodal artificial intelligence, focusing on addressing the diagnostic challenge in patients within the PSA gray zone. Specific objectives include: (1) improving screening efficiency to quickly identify high-risk individuals and avoid over-examination; (2) solving the diagnostic gray zone problem; (3) reducing unnecessary biopsies through non-invasive or minimally invasive precision tools; and (4) achieving personalized management through risk stratification. Study Design: Prospective enrollment of suspected prostate cancer patients. Total sample size is no less than 500 cases, divided into training set (approximately 400 cases) and validation set (approximately 100 cases) at an 8:2 ratio. Eligibility Criteria: Inclusion criteria: (1) age ≥45 years, male; (2) presenting with abnormal serum PSA (≥4 ng/mL), abnormal digital rectal examination, or suspicious lesions on prostate ultrasound; (3) undergoing prostate biopsy with definitive pathological results; (4) signed informed consent. Exclusion criteria: (1) previously diagnosed with prostate cancer and receiving surgery, radiotherapy, or endocrine therapy; (2) with other malignancies; (3) critical missing clinical data (e.g., missing PSA value, incomplete ultrasound report). Study Interventions/Assessments: All enrolled patients complete the following data collection: (1) serum PSA and free PSA; (2) routine laboratory tests including complete blood count, liver and kidney function; (3) transrectal or transabdominal prostate ultrasound with images stored in DICOM format and prostate volume recorded; (4) post-prostate massage urine for ctDNA methylation target detection; (5) digital rectal examination results, age, family history, medical history; (6) pathological diagnosis results from biopsy as gold standard. Models to be Developed: Tool 1 - ctDNA multimodal AI prediction model: using ctDNA methylation results combined with age, PSA, and prostate volume. Logistic regression and random forest will be compared. Tool 2 - Routine laboratory data-assisted decision model: integrating structured data including complete blood count, liver and kidney function, PSA, free PSA, age, and prostate volume. XGBoost and LightGBM with LASSO feature reduction will be used. Tool 3 - Prostate ultrasound image AI-assisted diagnostic model: using convolutional neural network (ResNet or DenseNet architecture) for deep learning modeling. The model outputs lesion probability heatmaps and malignancy probability scores. Multimodal Fusion Strategy: The three model outputs will be combined according to preset fusion rules to generate comprehensive risk stratification (low/moderate/high concern). Diagnostic sensitivity, specificity, positive predictive value, negative predictive value, and AUC of the fusion solution will be calculated using pathological results as gold standard. The AUC will be compared with that of PSA alone using DeLong test. Stratified analysis will be performed for the PSA 4-10 ng/mL gray zone subgroup. Decision curve analysis (DCA) will be used to evaluate clinical net benefit. Closed-loop Optimization: All pathological results will be periodically returned to the model management system in a de-identified manner, and quarterly iterative optimization of the three specialized models and fusion rules will be conducted. Study Duration: May 2026 to May 2028 (approximately 2 years). Funding: This is a hospital-level research project with an application fund of 50,000 RMB.