The answer is Yes. The cancer images eg: histopathology images are entirely different from coco dataset images like cat,, dog, van etc. As the coco dataset features are also different from cancer features. How the pretrained model features are fine tuning to learn new cancer features?
You could but a better alternative would be to use a network trained on the neuroimaging tasks like Chexnet!
From my experience (Mask R-CNN on chest CNCB-NCOV CT scans, pretrained on MS COCO), it's useful to use a pretrained backbone (e.g. ResNet50+FPN), but the RPN, RoI and Mask layers are all initialized from scratch. Pretrained backbone adds 2-3ppt accuracy.
@ Alex Ter-Sarkisov.... thank you... Even I too got good accuracy for cancer nuclei segmentation using mask r cnn pretrained on ms coco..
The thing is, I wanted to know how it works. Coco data is different in all aspects feom cancer cells.. but still I got good segmentation results... how? Any food for thoughts.
Pretrained backbone is useful because it extracts low-level features that are useful regardless of the problem at hand, e.g. geometric primitives. Also, MS COCO is a very diverse dataset, and therefore models pretrained on it will be able to detect a large number of those low-level features.
Yes... that's it... now I got... thank you so much.... please feel free to share more info regarding this if u have.. thank you
- Badges
- Science topic