riblidezso/frcnn_cad

Computer aided detection using Faster-RCNN

frcnn_cad's Issues

Why use CoreML?

I'm new to Osirix plugin development.. Is there a reason that you converted the Caffe model to CoreML and simplified? Was this just for speed or were you unable to call the Caffe model through the plugin?

Thanks!

reproducing the results on inbreast dataset

Hi, thanks for your great code. i tried to reproduce your result on inBreast dataset nut i couldn't get any similar or even close results. i tried to preprocess inbreast dicom images as you said in yor paper. is there any specific preprocessing that we should do? could you please add some code so we can regenerate your result on inbreast?

Downloading the Dataset

Hi! Would you be kind enough to let me know where can I download DREAM breast mammography image dataset? I cannot seem to find it. I am currently writing similar software, but using DDSM dataset which turns out to be smaller then I've expected. Thanks in advance!

Model license?

Dear Dezso,
Thanks for your inspiring work and wonderful model. We want to deploy your model for a commercial HPC cloud demo. The purpose of the demo is to assess inference speed of a medical AI model. Would you please provide license info if there is any?

Questions about INbreast ground truth

Dear Dezso,
Congratulations for the excellent work! It's amazing that your approach achieved so good results on DM challenge even without using the training data!
I have some questions about your paper and look for help.
First, what did you use as the ground truth for the INbreast data? I only found Bi-Rads scores in the meta data file, but it's not the real Benign/Malignant ground truth. In the INbreast paper it wrote: "A biopsy result for BI RADS 3, 4, 5, and 6 cases is also displayed whenever performed. The remaining cases were considered benign and therefore a biopsy was therefore not performed. Consequently, a biopsy was performed on 56 cases, of which 11 were found to be benign and the remaining 45 were malignant." However, I fail to identify which are those 45 malignant cases.
Second, the ROC curve on INbreast in your paper, was it measured per-image level or per-breast level? For example did you fuse the CC and MLO detection scores?
Finally, according to your paper and your achievement in DM challenge, faster-rcnn works amazingly well. However, I think faster rcnn actually ignores roi size information and changes the aspect ration because of the roi-pooling. These information might be important for cancer classification. Do you think there is a chance to do something in this aspect?
Thank you very much and looking for your reply!
Xuran

ignoring source layers

Hi,
Thanks to your grate work.
Have you ever seen this before ?

I0907 11:46:07.301913 23252 net.cpp:257] Network initialization done.
I0907 11:46:07.632989 23252 net.cpp:746] Ignoring source layer input-data
I0907 11:46:07.633008 23252 net.cpp:746] Ignoring source layer data_input-data_0_split
I0907 11:46:07.633029 23252 net.cpp:746] Ignoring source layer im_info_input-data_1_split
I0907 11:46:07.633031 23252 net.cpp:746] Ignoring source layer gt_boxes_input-data_2_split
I0907 11:46:07.642974 23252 net.cpp:746] Ignoring source layer rpn_cls_score_rpn_cls_score_0_split
I0907 11:46:07.643013 23252 net.cpp:746] Ignoring source layer rpn_bbox_pred_rpn_bbox_pred_0_split
I0907 11:46:07.643016 23252 net.cpp:746] Ignoring source layer rpn_cls_score_reshape_rpn_cls_score_reshape_0_split
I0907 11:46:07.643020 23252 net.cpp:746] Ignoring source layer rpn-data
I0907 11:46:07.643041 23252 net.cpp:746] Ignoring source layer rpn_loss_cls
I0907 11:46:07.643045 23252 net.cpp:746] Ignoring source layer rpn_loss_bbox
I0907 11:46:07.643049 23252 net.cpp:746] Ignoring source layer roi-data
I0907 11:46:07.713012 23252 net.cpp:746] Ignoring source layer loss_cls
I0907 11:46:07.713027 23252 net.cpp:746] Ignoring source layer loss_bbox

I dont know what is the problem ?
I will be happy if you help to sovle my issue.
Thanls a lot.

DDSM LJPEG optical density calibration

Dear Dezso,
I want to train a neural network model on the DDSM dataset. When I do optical density calibration, I think I get wrong results. I posted my python function which should be obvious to you and two samples( C_0029_1.LEFT_CC.jpeg, C_0217_1.RIGHT_MLO.jpeg) below.

Could you spare some time on this?

Thanks very much!

We have converted the lossless jpeg images to png format, mapped the pixel values to optical density using calibration functions from the DDSM website, and rescaled the pixel values to the 0–255 range.

`def optical_density_adjustment(image, case_dict):

if case_dict["source_label"] == "A" and case_dict["machine"] == "DBA":
    image = ( np.log10(image) - 4.80662 ) / (-1.07553)
elif case_dict["source_label"] == "A" and case_dict["machine"] == "HOWTEK":
     image = 3.789 - 0.00094568 * image
elif case_dict["source_label"] == "B" or case_dict["source_label"] == "C" and case_dict["machine"] == "LUMISYS":
     image = 3.6 - (image - 495) / 1000
elif case_dict["source_label"] == "D" and case_dict["machine"] == "HOWTEK":
     image = 3.96604095240593 + (-0.00099055807612) * image
else:
     print("ERROR: undefined machine type")
     exit()
image = cv2.normalize(image, None, 0, 255, cv2.NORM_MINMAX)
image = np.uint8(image)
return image

`

Developing a plugin for OsiriX

Hello,
Congrats on your work, it's really interesting and enriching. I really would like to reproduce your work.
I have a question regarding the OsiriX implementation, so you have developed a Caffe model then converted it to CoreML and afterward create the plugin for OsiriX.
Do you have open source your XCode project or do you have any tutorial to-recreate a plugin like you did ?
Thank you very much in advance,
Regards.