Computer & Information Science & Engineering, University of Florida, Gainesville, Florida, USA1
We present an ensemble of image-only convolutions neural network (CNN) models with different backbones and input sizes along with a self-supervised model to classify skin lesions. We have devised the first ensemble based on the winning solution to Kaggle's SIIM-ISIC Melanoma Classification challenge (We will be referring to this as the Ha-CNN model going further) and Bootstrap your own latent (BYOL) model which is based on self-supervised learning. The models have experimented with the SIIM-ISIC Melanoma dataset (2018-2020). Using specificity and sensitivity as the performance metrics, nine top-performing models were selected out of the eighteen models proposed in the Ha-CNN paper. We experimented BYOL model with two different backbones - ResNet and EfficientNet. The Ha-CNN model achieves a specificity and sensitivity of 94.3% and 92.1% with a negative predictive value of 99.2%. As with the BYOL model, our results show an increase of 1.00% for the ResNet-101 model supervision (94.73% and 93.40%) and an increase of 1.00% for the Efficient-B5 model (97.24% and 96.34%) with and without BYOL-self-supervision.
Two deep learning techniques are used to perform the melanoma classification. Each of them is described separately in the ./src/ folder. The CNN-Ensemble source code along with the repository setup is described in the ./src/CNN-Ha/ folder and the BYOL-CNN is described in the .src/BYOL/ Furthermore, the models results are saved in the ./results/ folder with different technique names.
The following table shows the comparison of different techniques on the task of Melanoma Classification:
| Technique name | Sensitivity (%) | Specificity (MCRMSE) | G-means | NPV | AUC |
|---|---|---|---|---|---|
| Dermatologists (Haenssle) | 88.9 ± 9.6 | 75.7 ± 11.7 | 82.0 | - | 82.0 |
| EGIR | 88 | 100 | 93.8 | > 99.0 | 95.5 |
| CNN-Thissen | 78 | 80 | 78.9 | - | - |
| CNN-Haenssle | - | 82.5 (@88.9 sen.) | 85.6 | - | 95.3 |
| CNN-Ha (paper) | - | - | - | - | 94.9 |
| CNN-Ha (experiments)* | 92.10 | 94.74 | 93.823 | 99.3 | 98.2 |
| BYOL-ResNet-101* | 42.81 | 97.94 | 64.75 | 94.73 | - |
| ResNet-101* | 27.05 | 98.44 | 51.6 | 93.40 | - |
| BYOL-EfficientNet-B5* | 59.16 | 98.44 | 76.3 | 96.24 | - |
| EfficientNet-B5* | 60.35 | 98.40 | 77.1 | 96.34 | - |
*The deep learning techniques show the results where we selected the threshold by maximizing the G-Means value. More details on the experiments and their results can be found in the results folder for the experiments CNN-Ensemble and the BYOL-CNN experiment.