search

Holliemin9090 / Mammographic-mass-CAD-via-pseudo-color-mammogram-and-Mask-R-CNN

Fully automatic computer-aided mass detection and segmentation via pseudo-color mammograms and Mask R-CNN
http://arxiv.org/abs/1906.12118
Other
24 stars 9 forks source link

reproducing the results on inbreast dataset #3

Closed CarrotC closed 4 years ago

CarrotC commented 4 years ago

Hi, thanks for your great code. I tried to reproduce your result on INbreast dataset but I couldn't get any similar or even close results. Having run Preprocess_mammo_mask.py, I got 2292 mask images and 299 pseudo color images because I deleted the 44 ones with invalid mask. After generating pseudo color image, I used the last 20% of the images for test with the trained model you uploaded. But I got the result Sensitivity:0.058524, False positive per image:0.65, Dice:0.85299@0.15935. I got images from DICOM files by code in https://github.com/danishm/mritopng. And I used code in https://gist.github.com/jendelel/3a8e768a8eb9345d49f2a82d02946122 to process xml files for raw annotation images. Is there any other specific preprocessing that we should do?

Holliemin9090 commented 4 years ago

Hi, thank you for your interest in the project. The “Preprocess_mammo_mask.py” should not generate that many masks. There are 410 INbreast images and 418 annotation masks (each one contains one lesion annotation or no lesion) after "Preprocess_mammo_mask.py" in our experiment. We did not have that many invalid masks to delete either. Maybe you could check if the mammograms and annotation masks generated from the xml files are correct. Since I received the mammograms and masks already in png format, I am not sure what could go wrong in transforming the data from xml to png.

CarrotC commented 4 years ago

@Holliemin9090 Sorry to disturb again. That's my fault in generating masks from xml files last time. But this time I do get 410 INbreast images and 418 annotation masks (each one contains one lesion annotation or no lesion) after "Preprocess_mammo_mask.py", and get bad results such as "Sensitivity:0.64286，False positive per image:2. 3049，Dice:0.89429@0. 040849". There are 10 models after running mamo_CAD.py train. I tested two of them and neither can get a good result.

Holliemin9090 commented 4 years ago

Hi, The training process is designed to save the model after each epoch so you will have 10 models for 10 epochs. The data were split into 60% training, 20% validation and 20% testing. We selected the model which had the lowest training+validation loss as the final model. The loss can be exported via tensorboard.