A recent Radiology journal study assesses the power of a fully automated deep learning (DL) model to produce deterministic outputs for identifying clinically significant prostate cancer (csPCa). Study: Fully Automated Deep Learning Model to Detect Clinically Significant Prostate Cancer at MRI. Image Credit: Antonio Marca / Shutterstock.
com Using machine learning to diagnose prostate cancer Prostate cancer is the second most common cancer affecting men throughout the world. To diagnose csPCa, multiparametric magnetic resonance imaging (MRI) is commonly used. A standardized reporting and interpretation approach involves the use of the prostate imaging reporting and data system (PI-RADS), which requires a high level of expertise.
Nevertheless, using PI-RADS to classify lesions is susceptible to intra- and inter-observer variation. Classic machine learning or DL can be used to detect csPCa by training a model on specific regions of interest that are informed by MRI scans. An alternative approach is to obtain predictions for each voxel by training a segmentation model.
These machine-learning approaches require a radiologist or pathologist to annotate the lesions at the model development stage, as well as the retraining and re-evaluation stages after clinical implementation. As a result, implementing these approaches is associated with high costs that also limit the data set's size. About the study The researchers of the current study were interested in developing a DL model to predict the presence of csPCa without prior information on the tumor's location.
They utilized patient-level labels clarifying the presence or absence of csPCa and compared the model's predictions with radiologists' predictions. Data were collected on patients without known csPCa who underwent an MRI scan between January 2017 and December 2019. T1-weighted contrast-enhanced images, T2-weighted images, apparent diffusion coefficient maps, and diffusion-weighted images were used to train a convolutional neural network to predict csPCa.
Pathologic diagnosis formed the reference standard. Four models were evaluated: image-only, radiologists, image + radiologist, and image + clinical + radiologist models. Related Stories New nanoparticle-based immunotherapy delays hormone therapy resistance in prostate cancer Rising cancer rates in younger generations signal urgent need for early-life prevention strategies Cannabis dependence associated with increased occurrence of head and neck cancer Four radiologists' PI-RADS ratings informed the external (ProstateX) test set and were used for the internal test set.
The DeLong test and receiver operating characteristic curves (AUCs) were used to evaluate radiologist performance. The tumor localization was shown using gradient-weighted class activation maps (Grad-CAMs). Study findings The image + clinical + radiologist model was associated with the highest predictive power with an AUC of 0.
94, followed by the image + clinical model with an AUC of 0.91. The image-only model and radiologists had an AUC of 0.
89. For the subset of pathologically proven cases within the internal set, the image + clinical model had the highest AUC at 0.88.
The radiologist model had an AUC of 0.78, whereas the clinical benchmark was associated with an AUC of 0.77.
Thus, the image + clinical + radiologist model had the highest predictive power among the entire internal test sample. In contrast, the image + clinical model had the highest predictive power in the subset of pathology-proven cases. For the image + clinical + radiologist model, the true-positive rate (TPR) was the highest, and the false-positive rate (FPR) was the lowest.
For pathologically proven cases, the radiologist's TPR was the highest, and the image + clinical model's FPR was the lowest. For the external dataset, the image + radiologist model showed the highest AUC and TPR and the lowest FPR. Concerning the use of Grad-CAM for tumor localization, patients with PI-RADS 1 or 2 lesions who did not undergo biopsy constituted a significant fraction of negative cases.
Several cases were labeled as false-negative. Conclusions The current study successfully predicted the presence of csPCa with MRI using a DL model. No statistically significant differences were observed between the model performance and that of experienced radiologists for both internal and external test sets.
These findings indicate that the DL model developed in the current study has the potential to assist radiologists in identifying csPCa and lesion biopsy, which could significantly improve prostate cancer diagnosis. A vital limitation of the current study is its single-site and retrospective nature. Furthermore, in an effort to improve its predictive accuracy, the DL model included only radiologists who specialized in prostate MRI and excluded trainees and general radiologists.
Cai, C. J., Nakai, H.
, Kuanar, S., et al . (2024) Fully Automated Deep Learning Model to Detect Clinically Significant Prostate Cancer at MRI.
Radiology 312 (2):e232635 doi:10.1148/radiol.232635.
