Deep Learning Classification of Prostate Cancer Using MRI Histopathologic Data.

Purpose To evaluate the diagnostic capability of MR histopathology (MRH) for identifying prostate cancer and guiding selection of imaging parameters for clinical MRH acquisition. Materials and Methods This retrospective study was conducted on a dataset of histologic slides of radical prostatectomy specimens with prostate cancer collected between 2009 and 2011. The dataset was used to perform an in silico validation of MRH, a method of assessing tissue texture that trades spatial coherence for spatial resolution surpassing traditional MRH by over an order of magnitude. The MRH measurement process was computationally recreated on the annotated slides, creating a dataset of spectral intensities at submillimeter wavelengths. Novel artificial intelligence analytics methods were developed to classify spectral data as normal or containing prostate cancer, and diagnostically informative parameters were determined. Results A set of spatial frequencies that maximized discriminative ability between healthy and cancerous tissue (area under the receiver operating characteristic curve, 0.79) was identified, thereby informing future clinical implementation. Integrating spatial context into the model denoised the inferential results and improved classification performance (area under the receiver operating characteristic curve, 0.84) and moreover enabled the estimation of lesion size. Conclusion This study demonstrates the feasibility of MRH as a novel method for prostate cancer detection and identifies imaging parameters that may guide clinical implementation. MRI, Prostate Cancer, Neural Networks, Histopathology, MR-Spectroscopy, Prostate, Tissue Characterization, Technology Assessment © RSNA, 2025 See also commentary by Fields and Hassan in this issue.