A custormized ResNet model for OCT colorectal image classification.
A custormized residual neural network (ResNet) was manufactured and trained to perform automatic image processing and real-time diagnosis of the OCT images.
The code was tested with Python 3.8.8 and TensorFlow 2.4.1.
Required packages:
- tensorflow-gpu 2.4.1
- numpy 1.20.1
- pandas 1.2.4
- matplotlib 3.3.4
- pillow 8.2.0
On Whitaker 160 ORIGIN GPU PC, use the enviroment OCT NN: conda activate OCTNN.
Use train.py to train the ResNet. The training dataset and validation dataset contains both bechtop and catheter images. They were cropped from B-scan images to 125 x 512 or 279 x 512.
The pre-trained networks can be found in the folder 'models/models_train_on_both2'. Models for all epochs were saved. The model with the lowest validation loss is save_at_103.h5 with an epoch of 103 and was used in the paper.
The testing dataset is in the 'whole_images2' folder and ended with '_test'.
Use plot_prob.py to plot prediction scores on all testing images and save the scores to csv files.
Use test_result.py to average over B-scan images, plot ROC, and calculate AUC.
Luo H, Li S, Zeng Y, Cheema H, Otegbeye E, Ahmed S, Chapman WC Jr, Mutch M, Zhou C, Zhu Q. Human colorectal cancer tissue assessment using optical coherence tomography catheter and deep learning. J Biophotonics. 2022 Feb 11:e202100349. doi: 10.1002/jbio.202100349. Epub ahead of print. PMID: 35150067.