We validated our previously developed algorithm for prostate cancer detection using an independent patient cohort who went through radical prostatectomy and compared with genitourinary radiologists under the same setting.
We developed a rapid, stage-free phase tomography by hand spinning the defocus knob while updating illumination patterns and taking measurements at a high frame rate, and the defocus trajectory can be inferred with the algorithm.
We proposed a multi-class CNN to jointly detect prostate cancer lesions and characterizes their histopathological aggressiveness by fully utilizing distinctive knowledge from multi-parametric MRI.
We proposed to automatically discover object parts in an unsupervised manner, which disentangles feature components of object parts from feature representations of each convolutional filter.
We used active question-answering to weakly-supervised semanticize neural patterns in conv-layers of the CNN and mine part concepts.