Ensemble of Convolutional Neural Networks for Disease Classification of Skin Lesions
Problem Statement: Fully supervised binary classification of skin lesions from dermatoscopic images.
Note: The following approach won 1st place in the 2019 Computer-Aided Diagnosis: Deep Learning in Dermascopy Challenge at Universitat de Girona scoring 92.2% accuracy (kappa: 0.819) at test-time, during the 2018-20 Joint Master of Science in Medical Imaging and Applications (MaIA) program.
Acknowledgments: Pavel Yakubovskiy for the TensorFlow.Keras implementation of EfficientNet, SEResNeXt-101 and SENet-154, and Mina Sami for the Python implementation of Shades of Gray Color Constancy.
Data: Class A: Nevus; Class B: Other (Melanoma, Dermatofibroma, Pigmented Bowen's, Basal Cell Carcinoma, Vascular, Pigmented Benign Keratoses) [4800/1200/1000 : Train/Val/Test Ratio]
Directories
● Preprocessing Pipeline for Color Space/Constancy: scripts/color-io.ipynb
● Individual Model Training-Validation Pipeline: scripts/train-val.ipynb
● Ensemble Validation Pipeline: scripts/ensemble-val.ipynb
● Ensemble Inference Pipeline: scripts/ensemble-test.ipynb
Train/Test-Time Data Augmentation
Figure 1. All 5 different types of data augmentation [vertical (b)/horizontal (c) flips, brightness shift (d), saturation (e)/contrast (f) boost) used at train-time to broaden the data representation beyond limited pre-existing samples, and test-time to ensure a full prediction from the classifier that is unaffected by the orientation or lighting conditions of the scan. Predictions from all 6 variations [including the original (a)] are averaged to obtain the final prediction per sample.
Multi-Scale Input
Figure 2. Original RGB image (left), center cropped 448 x 448 x 3 image used to train 3 CNN member models and the further center cropped 224 x 224 x 3 image used to train 2 more CNN member models. Each model learns to classify at a different scale, with the hypothesis that the collective ensemble benefits from a multi-scale input.
Feature Maps
Figure 3. Features maps derived from the output of the second block of expanded convolutional layers in a pre-trained EfficientNet-B6 with ImageNet weights, after passing an input skin lesion image through the network.
Figure 4. Features maps derived from the output of the second block of expanded convolutional layers in a finetuned EfficientNet-B6 initialized with ImageNet weights, after passing an input skin lesion image through the network.
Experimental Results
Figure 5. Validation performance for the collective ensemble and each member model. Accuracy, sensitivity and specificity scores are calculated at the default threshold of 0.50.
Gradient Class Activation Maps
Figure 6. Gradient–Class Activation Maps (Grad-CAM) from finetuned EfficientNet-B6 –using the gradients of the nevus class flowing into the final convolutional layer, to produce a coarse localization map highlighting important regions in the image for predicting nevus.