Skip to content
This repository has been archived by the owner on Sep 1, 2021. It is now read-only.

Latest commit

 

History

History
50 lines (37 loc) · 3.26 KB

README.md

File metadata and controls

50 lines (37 loc) · 3.26 KB

Skin Lesion Analysis Towards Melanoma Detection

Skin cancer is the most common cancer globally, with melanoma being the most deadly form. Dermoscopy is a skin imaging modality that has demonstrated improvement for diagnosis of skin cancer compared to unaided visual inspection. In order to make expertise more widely available, the International Skin Imaging Collaboration (ISIC) has developed the ISIC Archive, an international repository of dermoscopic images, for both the purposes of clinical training, and for supporting technical research toward automated algorithmic analysis by hosting the ISIC Challenges.

The goal for the ISIC 2019 challenge is to classify dermoscopic images among nine different diagnostic categories:

Melanoma (MEL)
Melanocytic nevus (NV)
Basal cell carcinoma (BCC)
Actinic keratosis (AK)
Benign keratosis (solar lentigo / seborrheic keratosis / lichen planus-like keratosis) (BKL)
Dermatofibroma (DF)
Vascular lesion (VASC)
Squamous cell carcinoma (SCC)
None of the others (UNK)

25,331 images are available for training across 8 different categories. Additionally, the test dataset contains an additional outlier class not represented in the training data, which developed systems must be able to identify.

Manuscript

We documented our approach to the ISIC-2019 challenge at https://arxiv.org/abs/2101.03814

Training Data

Combined data from ISIC-2019, PH2, Light Field Image Dataset of Skin Lesions, SD-198, 7-point criteria evaluation Database, MED-NODE.

Web App

The best single model has been deployed as a web app at https://skin-lesion-classifier-285410.uc.r.appspot.com/

Accuracy

An ensemble of classifiers performs at 0.634 balanced multiclass accuracy: ISIC 2019 Live Leaderboard (2019: Lesion Diagnosis).

Source

Ensembling (ensemble.ipynb) has been achieved by calculating the arithmetic mean of the predictions of the best performing models. Trained model architectures are

If GitHub is unable to render the notebooks they can be viewed externally at https://nbviewer.jupyter.org/github/j05t/lesion-analysis/tree/master/.

Setup

pip install fastai==1.0.61 efficientnet-pytorch==0.7.0 pretrainedmodels==0.7.4 \
            jupyter==1.0.0 matplotlib==3.3.2 pandas==1.1.4 seaborn==0.11.0 ipython==7.19.0