MP18-13: Transformer-based large model with mp-MRI for non-invasive auxiliary diagnosis of prostate cancer: a multi-center study
Friday, May 3, 2024 3:30 PM to 5:30 PM · 2 hr. (US/Central)
302B
Abstract
Information
Full Abstract and Figures
Author Block
Chao Liang*, Meiling Bao, Nanjing, China, Ye Yan, Zhuhong Shao, Beijing, China, Jie Li, Nanjing, China
Introduction
The non-invasive diagnosis of prostate cancer before surgery still faces challenges and a growing need for precision diagnosis and treatment. Discordance between clinical tools (e.g. PSA, PI-RADS) and pathology leads to over-diagnosis or under-diagnosis. Transformer-based large model provides a more efficient way to build cross-scale information mapping. The study first utilizes a transformer model to analyze the correlation between mp-MRI and pathological characterization. The study aimed to provide non-invasive auxiliary diagnostic models for clinics and validate in multi-center.
Methods
Our study retrospectively collected 3008 patients who underwent pathological evaluation of the prostate from January 2013 to October 2022, which 1788 patients were from Peking University Third Hospital and 1210 patients were from The First Affiliated Hospital of Nanjing Medical University. The study proposed an MRI-based Predicted Transformer for Prostate cancer (MRI-PTPCa) to translate mp-MRI to prostate tumor aggressiveness, and use the MRI-PTPCa to predict grade group for prostate cancer. The MRI-PTPCa was trained by more than 160,000 iterations and variant training samples greater than 100,000, which were augmented from 1748 patients’ mp-MRI of the training cohort. Finally, the performance of MRI-PTPCa was evaluated on more than 1,000 multi-center patients.
Results
The MRI-PTPCa for diagnostic tasks in prostate cancer earned performance with both accuracy and robustness. The performance of MRI-PTPCa for non-invasive diagnosis of prostate cancer yield area under the curve (AUC) of 0.981 [0.976-0.986], sensitivity (SEN) of 0.989 [0.986-0.993] and specificity (SPE) of 0.97 [0.96-0.98] in external testing. The re-diagnosis benefits of the MRI-PTPCa in the PSA gray zone have been improved, which also yielded better net benefit than clinical radiology evaluation by PI-RADS. For diagnosing clinically significant prostate cancer, MRI-PTPCa once again achieved better accuracy than the clinical method. The MRI-PTPCa yielded an overall AUC of 0.978 [0.975-0.981] for clinically significant prostate cancer diagnosis, which earned SEN of 0.998 [0.996-0.999] and SPE of 0.885 [0.867-0.902].
Conclusions
MRI-PTPCa takes advantage of mp-MRI and retrospective big data. Correlation learning between imaging and pathology helps to obtain stronger models. It demonstrates better and more robust performance in the diagnosis of prostate cancer and clinically significant prostate cancer diagnosis.
Source Of Funding
National Natural Science Foundation of China (grant number 82002718)
Author Block
Chao Liang*, Meiling Bao, Nanjing, China, Ye Yan, Zhuhong Shao, Beijing, China, Jie Li, Nanjing, China
Introduction
The non-invasive diagnosis of prostate cancer before surgery still faces challenges and a growing need for precision diagnosis and treatment. Discordance between clinical tools (e.g. PSA, PI-RADS) and pathology leads to over-diagnosis or under-diagnosis. Transformer-based large model provides a more efficient way to build cross-scale information mapping. The study first utilizes a transformer model to analyze the correlation between mp-MRI and pathological characterization. The study aimed to provide non-invasive auxiliary diagnostic models for clinics and validate in multi-center.
Methods
Our study retrospectively collected 3008 patients who underwent pathological evaluation of the prostate from January 2013 to October 2022, which 1788 patients were from Peking University Third Hospital and 1210 patients were from The First Affiliated Hospital of Nanjing Medical University. The study proposed an MRI-based Predicted Transformer for Prostate cancer (MRI-PTPCa) to translate mp-MRI to prostate tumor aggressiveness, and use the MRI-PTPCa to predict grade group for prostate cancer. The MRI-PTPCa was trained by more than 160,000 iterations and variant training samples greater than 100,000, which were augmented from 1748 patients’ mp-MRI of the training cohort. Finally, the performance of MRI-PTPCa was evaluated on more than 1,000 multi-center patients.
Results
The MRI-PTPCa for diagnostic tasks in prostate cancer earned performance with both accuracy and robustness. The performance of MRI-PTPCa for non-invasive diagnosis of prostate cancer yield area under the curve (AUC) of 0.981 [0.976-0.986], sensitivity (SEN) of 0.989 [0.986-0.993] and specificity (SPE) of 0.97 [0.96-0.98] in external testing. The re-diagnosis benefits of the MRI-PTPCa in the PSA gray zone have been improved, which also yielded better net benefit than clinical radiology evaluation by PI-RADS. For diagnosing clinically significant prostate cancer, MRI-PTPCa once again achieved better accuracy than the clinical method. The MRI-PTPCa yielded an overall AUC of 0.978 [0.975-0.981] for clinically significant prostate cancer diagnosis, which earned SEN of 0.998 [0.996-0.999] and SPE of 0.885 [0.867-0.902].
Conclusions
MRI-PTPCa takes advantage of mp-MRI and retrospective big data. Correlation learning between imaging and pathology helps to obtain stronger models. It demonstrates better and more robust performance in the diagnosis of prostate cancer and clinically significant prostate cancer diagnosis.
Source Of Funding
National Natural Science Foundation of China (grant number 82002718)
Sessions
MP18: Imaging/Uroradiology I
302B