/MF-TAPNet

[MICCAI'19] Incorporating Temporal Prior from Motion Flow for Instrument Segmentation in Minimally Invasive Surgery Video

Primary LanguagePython

first unzip the dataset, and move each dataset (1-8) to ./data/train

Problems:

  1. how to deal with attention maps when training paused and restored?
  2. Sigmoid and Softmax in Attention Module (Not trained)
  3. bn in train and valid
  4. (IMPORTANT) whether using softmax operation in attention module or not
  5. (IMPORTANT)use + or * or conv2d(kernel 1) in final attention module operation

Handle:

  1. Dice in binary classification
  2. metrics for multiclass classification (Not Debuged)
  3. larger Attention map
  4. multiple config.py (for preprocess_data.py)
  5. Semi-supervised model

TRAIN_RECORD for each model: Binary:

  1. UNet (not resized) batchsize=8, learn_rate=1e-5, epoches=20, gpu=4 (TITAN V), validation IoU=0.6728385997379875: validation Dice=0.784111333214352
  2. UNet11 (not resized) batchsize=8, learn_rate=1e-5, epoches=20, gpu=4 (TITAN V), validation IoU=0.8175544938356124: validation Dice=0.8885956894127561
  3. UNet16 (not resized) batchsize=8, learn_rate=1e-5, epoches=20, gpu=4 (TITAN V), validation IoU=: validation Dice=