研究目的
To propose a deep transfer learning (DTL)-based model to distinguish indolent from clinically significant prostate cancer (PCa) lesions and to compare the DTL-based model with a deep learning (DL) model without transfer learning and PIRADS v2 score on 3 Tesla multi-parametric MRI (3T mp-MRI) with whole-mount histopathology (WMHP) validation.
研究成果
The proposed DTL-based model outperformed the DL-based model without transfer learning, confirming the contribution of transfer learning. The DTL-based model performance generated comparable performance to the expert reader PIRADS v2 score (p = 0.89), showing great potential to augment PCa for non-experts. This model would need to be validated in much larger datasets to further evaluate its clinical utility.
研究不足
One limitation of this study is the small sample size for testing because of the limited available labeled data. Another limitation is that we included a manual segmentation of the prostate to assist the normalization for the cases with the endorectal coil. The system requires the lesion detection as the input to define an image patch.
1：Experimental Design and Method Selection
The DTL-based model was trained on 169 lesions in 110 arbitrarily selected patients and tested on the remaining 47 lesions in 30 patients. The model was compared with the same DL model architecture trained from scratch and the classification based on PIRADS v2 score with a threshold of 4 using accuracy, sensitivity, specificity, and area under curve (AUC).
2：Sample Selection and Data Sources
140 patients with 3T mp-MRI and WMHP comprised the study cohort. The ground-truth was lesions detected by genitourinary (GU) pathologist on post robotic-assisted laparoscopic prostatectomy WMHP, blinded to all MRI information.
3：List of Experimental Equipment and Materials
3T mp-MRI was performed on a variety of scanners (Trio, Verio, Prisma or Skyra, Siemens Healthineers, Erlangen Germany) using a pelvic phase-array coil with or without the endorectal coil. Each scan used a standard mp-MRI scanning protocol including 3D axial T2 images using SPACE sequence, echo-planar imaging DWI sequence, and DCE images using TWIST.
4：Experimental Procedures and Operational Workflow
The DTL-based model was trained on the augmented training set and evaluated on the testing set. The input data were T2 SPACE and ADC images with each lesion contoured on both sequences using OsiriX. After proper pre-processing, the image patches enclosing the lesion were generated as the input to the proposed DTL based model.
5：Data Analysis Methods
The classification performance was quantitatively evaluated by accuracy, sensitivity, specificity, and area under curve (AUC) of receiver operating characteristics (ROC) curve using 47 testing lesions from 30 cases. Bootstrapping with 2000 resamples was performed to estimate the 95% confidence interval (CI) for AUC and DeLong test was used to compare the AUC of the DTL-based model and other models as well as PIRADS v2 score.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
