Nature Cancer：任善成团队等开发AI大模型，实现前列腺癌无创精准诊断与分级

Core Insights - Prostate cancer is the second most common cancer among men globally, with a rapid annual increase in incidence in China at 13%, now ranking sixth among male malignancies [2] - The number of new prostate cancer cases in China is projected to reach 144,000 in 2024, 199,000 by 2030, and 250,000 by 2035 [2] Diagnosis Challenges - Diagnosis primarily relies on PSA blood tests, ultrasound, and digital rectal exams, with 1/3 of men over 50 showing suspicious nodules and nearly 10% having elevated PSA levels [3] - The PI-RADS scoring system for MRI has significant subjective and accuracy flaws, leading to potential misdiagnosis and unnecessary procedures [3] Need for Advanced Tools - There is an urgent need for an efficient, accurate, and non-invasive diagnostic tool to assist in the diagnosis and grading of clinically suspicious prostate cancer patients [4] - The emergence of AI technologies offers new possibilities for correlating imaging data with pathological results, paving the way for non-invasive diagnosis [4] AI Model Development - A multi-center study developed and validated an AI-based model, MRI-PTPCa, for efficient, accurate, and non-invasive diagnosis and grading of prostate cancer [5][11] - The model integrates advanced techniques such as self-supervised learning and transfer learning, significantly enhancing predictive performance [7] Model Performance - The MRI-PTPCa model demonstrated high consistency with pathological evaluations, outperforming clinical assessments and other predictive models, achieving an AUC of 0.983 for prostate cancer detection [9] - The model's predictive accuracy for grading was 89.1%, indicating its potential as a new non-invasive diagnostic tool [9] Interpretability and Validation - The study provided a comprehensive analysis correlating MRI-PTPCa scores with Gleason grading, highlighting the model's interpretability through visual heatmaps and quantitative features [10] - The model's features were significantly associated with various pathological characteristics, supporting the feasibility of linking imaging and pathology [10]