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

Figure 8

An example of model detection: a 70-year-old man in the testing set with prostate-specific antigen of 30.2 ng/mL and PI-RADs of 5. (a) The patient’s prostate is visualized in 3D coordinates for volume representation. (b) The patient’s prostate surface and its ROI (yellow) outlined by radiologists. (c) The patient’s prostate surface and its potential cancer distribution (dark yellow) generated by the WSUNet. (d) The patient’s original ROI (yellow) and each biopsy core (Gleason marked as red and <7 masked as green). (e) The ROI (dark yellow) detected by the model and each biopsy core (Gleason marked as red and <7 masked as green). (f) A slice of the T2-weighted prostate MRI. (g) Slices with the detected ROI (brilliant yellow) and the positive biopsy core (red). In this meaningful example, targeted biopsy yielded no meaningful results, and the model presented a cancer distribution, part of which was validated by systematic biopsy. In this example, the model successfully detected MIPCas, which radiologists have not outlined.

(a)

(b)

(c)

(d)

(e)

(f)

(g)