This is a Keras with TensorFlow backend implementation for ISIC 2019 Challenge Task 1: classify dermoscopic images among nine different diagnostic categories without meta-data. The proposed approaches implemented in this repository were ranked 25^th out of 64 unique teams with Balanced Multiclass Accuracy of 0.505 (0.636 for the first team). The ODIN (Out-of-DIstribution detector for Neural networks) method used in the first approach was re-implemented using Keras. This is also my capstone project of Udacity Machine Learning Nanodegree. For more details please refer to the project report.

• Python 3.5 or above
• Keras 2.2.4
• TensorFlow 1.14
• pandas
• NumPy
• Matplotlib
• scikit-learn
• OpenCV-Python
• tqdm
• Augmentor: A modification of Augmentor 0.2.3 is under Augmentor
• Pillow-SIMD and libjpeg-turbo: Not necessary but recommend using them to improve image processing performance. I followed the fastai's document to install them.

• ISIC 2019 Training/Test Data
• ISIC Archive

• Seven-Point Checklist Dermatology Dataset (Github)

After downloading the dataset, all melanosis images can be retrieved from the dataset by using the notebook.

• Put all data under a root folder like the tree structure shown below.
  • ISIC_2019_Training_Input and ISIC_2019_Test_Input folders contain images of ISIC 2019 Training Data and Test Data respectively.
  • Out_Distribution folder contains all out-of-distribution images from both ISIC Archive and Seven-Point datasets.
  • ISIC_2019_Training_GroundTruth_DuplicateRemoved.csv is a copy of ISIC_2019_Training_GroundTruth.csv but removing two entries: ISIC_0067980 and ISIC_0069013.

RootFolder
│   ISIC_2019_Training_GroundTruth.csv
│   ISIC_2019_Training_GroundTruth_DuplicateRemoved.csv
│
└───ISIC_2019_Test_Input
│   │   ISIC_0034321.jpg
│   │   ISIC_0034322.jpg
│   │   ...
│   │
└───ISIC_2019_Training_Input
│   │   ISIC_0000000.jpg
│   │   ISIC_0000001.jpg
│   │   ...
│   │
└───Out_Distribution
│   │   7pt_Fgl059.jpg
│   │   7pt_Fhl002.jpg
│   │   ...
│   │   ISIC_0001115.jpg
│   │   ISIC_0001129.jpg
│   │   ...

Follow the steps starting from the Common Parameters cell in the notebook approach_1.ipynb and approach_2.ipynb to reproduce training and predicting processes of approach-1 and 2 respectively. You might need to change a few parameters according to the environment.

Here is an example code of training DenseNet201, Xception and ResNeXt50 models using approach-1. Assuming the path to the data root folder is /home.

python3 main.py /home --approach 1 --training --epoch 100 --batchsize 32 --maxqueuesize 10 --model DenseNet201 Xception ResNeXt50

Use -h to show usage messages.

python3 main.py -h

Please visit ISIC 2019 Challenge Leaderboards for complete rankings.

Team Rankings

Ranked 25^th out of 64 unique teams with Balanced Multiclass Accuracy of 0.505 (0.636 for the first team)

Approach Rankings

MIT