TransNorm: Transformer Provides a Strong Spatial Normalization Mechanism for a Deep Segmentation Model
The official code for "TransNorm: Transformer Provides a Strong Spatial Normalization Mechanism for a Deep Segmentation Model".
R. Azad, Mohammad Al-Antary, Moein Heidari, and Dorit Merhof , "TransNorm: Transformer Provides a Strong Spatial Normalization Mechanism for a Deep Segmentation Model", download link.
- July 17, 2022: Initial release.
- July 14, 2022: Submitted to IEEE Access Journal [Under Review].
In this paper, we argue that combining the two descriptors, namely, CNN and Transformer might provide an efficient feature representation, which is at the heart of our research in this paper. Majority of existing CNN-Transformer based networks suffer from a weak construction on the skip connection section. To this end, we design a two-level attention mechanism based on the Transformer module to adaptively recalibrate the feature combination on the skip connection path.
- June 16, 2022: First release (Complete implemenation for SKin Lesion Segmentation on ISIC 2017, SKin Lesion Segmentation on ISIC 2018 and SKin Lesion Segmentation on PH2 dataset added.)
This code has been implemented in python language using Pytorch library and tested in ubuntu OS, though should be compatible with related environment. following Environement and Library needed to run the code:
- Python 3
- Pytorch
For training deep model and evaluating on each data set follow the bellow steps:
1- Download the ISIC 2018 train dataset from this link and extract both training dataset and ground truth folders inside the dataset_isic18.
2- Run Prepare_ISIC2018.py for data preperation and dividing data to train,validation and test sets.
3- Run train_skin.py for training the model using trainng and validation sets. The model will be train for 100 epochs and it will save the best weights for the valiation set.
4- For performance calculation and producing segmentation result, run evaluate_skin.py. It will represent performance measures and will saves related results in results folder.
Notice: For training and evaluating on ISIC 2017 and ph2 follow the bellow steps :
ISIC 2017- Download the ISIC 2017 train dataset from this link and extract both training dataset and ground truth folders inside the dataset_isic18\7.
then Run Prepare_ISIC2017.py for data preperation and dividing data to train,validation and test sets.
ph2- Download the ph2 dataset from this link and extract it then Run Prepare_ph2.py for data preperation and dividing data to train,validation and test sets.
Follow step 3 and 4 for model traing and performance estimation. For ph2 dataset you need to first train the model with ISIC 2017 data set and then fine-tune the trained model using ph2 dataset.
In bellow, results of the proposed approach illustrated.
In order to compare the proposed method with state of the art appraoches on SKin Lesion Segmentation, we considered Drive dataset.
| Methods (On ISIC 2017) | Dice-Score | Sensivity | Specificaty | Accuracy |
|---|---|---|---|---|
| Ronneberger and et. all U-net | 0.8159 | 0.8172 | 0.9680 | 0.9164 |
| Oktay et. all Attention U-net | 0.8082 | 0.7998 | 0.9776 | 0.9145 |
| Lei et. all DAGAN | 0.8425 | 0.8363 | 0.9716 | 0.9304 |
| Chen et. all TransU-net | 0.8123 | 0.8263 | 0.9577 | 0.9207 |
| Asadi et. all MCGU-Net | 0.8927 | 0.8502 | 0.9855 | 0.9570 |
| Valanarasu et. all MedT | 0.8037 | 0.8064 | 0.9546 | 0.9090 |
| Wu et. all FAT-Net | 0.8500 | 0.8392 | 0.9725 | 0.9326 |
| Azad et. all Proposed TransNorm | 0.8933 | 0.8532 | 0.9859 | 0.9582 |
For more results on ISIC 2018 and PH2 dataset, please refer to the paper
You can download the learned weights for each dataset in the following table.
| Dataset | Learned weights (Will be added) |
|---|---|
| ISIC 2018 | TransNorm |
| ISIC 2017 | TransNorm |
| Ph2 | TransNorm |
All implementations are done by Reza Azad and Moein Heidari. For any query please contact us for more information.
rezazad68@gmail.com
moeinheidari7829@gmail.com