Chainer Skin Lesion Detector

Skin Lesion Detector using HAM10000 dataset with Chainer

Requirements

• Python 3.6
• Chainer 5.0
• ChainerCV 0.11
• cupy-cuda90
• opencv 3.4

$ pip install -r requirements.txt

Setup environment using Docker

$ docker build -t chainer-skin-lesion-detector .
$ docker run --rm -it -v $(pwd):/chainer-skin-lesion-detector --runtime nvidia --name chainer-skin-lesion-detector-dev chainer-skin-lesion-detector /bin/bash

Directory architecture

.
├── data
│   ├── ISIC2018_Task1-2_Training_Input
│   ├── ISIC2018_Task1_Training_GroundTruth
│   ├── preprocessed
│   │   ├── ground_truth
│   │   └── input
│   └── xml
└── src
    ├── models
    │
    ├── notebooks
    │
    ├── result
    └── util

Download dataset

• Download training dataset and ground truth data from Task 1: Training | ISIC 2018 to data/ directory

Preprocess

• Re-scale image and ground truth
• Make bounding box from ground truth of segmentation image
• Create VOC format based label to data/xml directory

$ python make_dataset.py --loaderjob 4

An example of annotation data with a bounding box from the ground truth of segmentation using ISIC2018 task1 dataset:

Train

• You can specify model, number of batch size, number of epoch, GPU ID and number of parallel data loading process.

$ python main.py --model ssd300 --batchsize 32 --epoch 30 --gpu 0 --loaderjob 4

Evaluation

Example of model prediction

Reference

• Tschandl, Philipp, Cliff Rosendahl, and Harald Kittler. "The HAM10000 Dataset: A Large Collection of Multi-Source Dermatoscopic Images of Common Pigmented Skin Lesions." arXiv preprint arXiv:1803.10417 (2018).
• Liu, Wei and Anguelov, Dragomir and Erhan, Dumitru and Szegedy, Christian and Reed, Scott and Fu, Cheng-Yang and Berg, Alexander C. "SSD: Single shot multibox detector." 14th European Conference on Computer Vision, ECCV 2016. Springer Verlag, 2016.