Anoushkrit Goel:
Exceptions in the “WBerg” dataset which contain multiple sub-cases and have different BI-RADS score for respective sub-case. For example: WB129A and WB129B are the two cases under WB129 which have different BI-RADS score.
Proposed Solution: Sort all the anomalies like this and classify them under test data for the model because model will eventually classify the BI-RADS score of those unknown sub-datasets. Anomalies:
- WB102
- WB114
- WB120
- WB129
- WB137
- WB163
- WB202
- WB208
- WB245
- WB249
- WB252
- WB253
- WB254
- WB268
- WB269
- WB271
- WB272
- WB281
- WB282
- WB286
- WB288
- WB290
- WB297
- WB302
- WB305
- WB312
- WB323
- WB327
- WB338
- WB344
- WB349
- WB361
- WB369
- WB372
- WB379
- WB384
- WB387
- WB388
- WB398
- WB405
- WB415
- WB424
- WB430
Anoushkrit Goel:
- Creating a massive dataset of Ultrasound images which contain callipers and leaving those which do not.
- Palp8_31, Palp_8_6, PalpB9Clear226,
- Philips Ultrasound Scanning device is not good in creating contrasting images to detect lesions perfectly.
- Siemens cannot be used WB264 contains no callipers.
Anoushkrit Goel:
The model was initially destined to be trained on un-callipered images by creating bounding box around them using imglab. Coordinates of these bounding boxes would have been written in an .xml file by the same software which would be eventually fed to the model. But, we are not certain as to what type of images will be provided to the deployed model while, hence, the model needs to be trained with both type of images. Technically, all the type of images regardless of callipers or metadata written on the image itself will be used to train the following.
- Break His Dataset
- Discovered the Attention Gated Networks using PyTorch.