Detecting MRI-Invisible Prostate Cancers Using a Weakly Supervised Deep Learning Model

Figure 4

The workflow for material collection, model training, and validation. Materials included T2-weighted MRI, MRI location, Gleason grade of each biopsy, and the outline of the lesion and prostate, which were outlined by a radiologist (the delineation of lesions and the MRI-invisible appearance of MIPCas were both based on multimodal MRI), and clinical features, which include PI-RADS of the outlined lesions and the patient’s prostate-specific antigen. Training: the model was trained to classify whether the Gleason grade was greater than or equal to 7 for each biopsy core to obtain the ability to correlate spatial location with pathological information. Validation: the model generates cancer distribution maps for each MRI in the testing set, and these maps were validated by systematic biopsy.