U-Net is a renowned convolutional neural network (CNN) designed primarily for biomedical image segmentation.

• Symmetrical Design: Comprising an encoder (contracting path) and a decoder (expansive path).
• Skip Connections: Retain spatial context by bridging encoder and decoder.
• Efficient and Precise: Requires fewer training samples but offers high precision.

1. Encoder:

   • Two 3x3 convolutions followed by a ReLU and a 2x2 max pooling.
   • Feature channels double after each pooling.

2. Bottleneck:

   • Two 3x3 convolutions with ReLU activation.

3. Decoder:

   • Upsampling and a 2x2 transposed convolution.
   • Concatenation with skip-connected feature map from the encoder.
   • Two 3x3 convolutions with ReLU.

4. Output:

   • 1x1 convolution to map to the number of classes.

• Original: Biomedical image segmentation.
• Extended: Any Semantic segmentation tasks.

For a deeper dive, check out the link_to_paper).

We have implemented U-Net on the Binary instrument segmentation task from the ROBUST-MIS Challenge dataset.

We have used a binary segmentation dataset from the ROBUST-MIS Challenge. You can down the dataset from here.

We got the best results with DDANet on this dataset.

The official link to the DDANet code can be found here.

Please cite our work Exploring Deep Learning Methods for Real-Time Surgical Instrument Segmentation in Laparoscopy if you find it useful.