Beating Melanoma with Deep Learning: letting the data speak

The main focus of the reserach is to perform analysis of skin lesions with deep learning pipeline systems.

Getting Started

These instructions will get you a copy of the project up and running on your local machine.

Prerequisites

Pytorch 0.2.0
(Do not use Pytorch 0.3 or higher, the codes are not updated to work with pytorch 0.3 )

Hair removal

The hair removal is implemented using MATLAB.

Say what the step will be

1) Specify root directory for input images in main.m
2) Dermoscopic images can be super highresolution (7000x7000 or higher), resize images to get faster results.
Resizing might require suitable adjustment to radius of structuring element and threshold value in my_function.

Image Generation (DCGAN)

Training:

Fake images will be saved under images folder as fake_samples_ep%d_it%d.png after every 10th batch of real data. The model weights are saved as 'netG.pth' and 'netD.pth'.

usage main.py: Paste all the available data inside Generator/Data/Melanoma

Mean squared error check:

usage mse_check.py: find closest match between fake images and real images.

1) The images are saved under results folder. Real closest matched images are saved as image%d_m_%f.jpg%(idx, error).
2) Generated images are saved as image%d_m.jpg %(idx).

Lesion segmentation (Modified Unet)

Model weights can be downloaded from: https://drive.google.com/file/d/12c5W8EOzfPsiLVBoR6m7znbDPJEVLrA1/view?usp=sharing

Training

usage Unet.py: [--lr LEARNING_RATE] [--lr_de LEARNING_RATE_DECAY] [--checkpoint LOAD_PREV_CHECKPOINT] [-wd WEIGHT_DECAY] [-rd ROOT_DIR]

1) Checkpoints are saved at every epoch as checkpoint_ep%d.pth.tar %(epoch)
2) Validation and Training losses are saved at every epoch as train_loss%d.pkl %(epoch)
and val_loss%d.pkl %(epoch).
3) Output images are save under results folder as:
    i)  INPUT_IMAGE: val_epoch%ditr%d_a.jpg %(epoch,iteration)
    ii) LABEL: val_epoch%ditr%d_b.jpg %(epoch,iteration)
    iii) OUTPUT_MAP: val_epoch%ditr%d_c.jpg %(epoch,iteration)

Testing

usage testing.py

1) Specify model weights in the python script testing.py.
1) The value of average Jaccard Index is printed on the terminal.
2) Output images are save under results folder as:
    i)  INPUT_IMAGE: val_epoch%ditr%d_a.jpg %(epoch,iteration)
    ii) LABEL: val_epoch%ditr%d_b.jpg %(epoch,iteration)
    iii) OUTPUT_MAP: val_epoch%ditr%d_c.jpg %(epoch,iteration)

Classification (Pretrained ResNet 50)

The weights for the Classification model are available on: https://drive.google.com/file/d/12XZrDWHODCpfCB6Vh9TQdFKHP51ORcKp/view?usp=sharing

The data loader class (data_loader.py) take csv file as input. The CSV file is arranged as follows:

Image_ID, Label
ISIC_0015284, 0
ISIC_0015284, 0
.
.
.

Training

usage Train.py [--lr LEARNING_RATE] [--lr_de LEARNING_RATE_DECAY]
[--checkpoint LOAD_PREV_CHECKPOINT] [-wd WEIGHT_DECAY] [-rd ROOT_DIR]

1) Checkpoints are saved at every epoch as checkpoint_ep%d.pth.tar %(epoch)
2) A list of loss and accuaracy values are saved as train_loss.pkl | val_loss.pkl
and train_corrects.pkl | val_corrects.pkl

Testing

usage Test.py

1) The scores for each images is saved in a numpy file as submission.npy
2) Run plot_roc.py to generate various plots.

Bonus

API to download segmentation masks of images obtained from ISIC-Archive is available in the folder Bonus.

Acknowledgments

We thank Dr. Marius Bojarski NVIDIA Corporation for inpirations and useful feedbacks.