Skin Carcinoma Detection Using Parallel Deep Residual Networks

A deep neural network developed following the residual learning and separable convolution paradigms to diagnose basal and squamous cell carcinoma using a subset of ISIC dataset. The network has been designed using the Keras framework.

Dataset used for model development

The dataset used for this project can be found within the "data_resized" folder in compressed format. This dataset is a subset of the ISIC archive data.

Data preprocessing

The image data has already been resized to 224 x 224 dimensions. However, we apply further preprocessing steps which can be implemented using the "preprocessing.py" file.

Proposed network

The deep neural network devised by us is defined in the "carcinomaNetwork.py" file. Please use the compile_model() method to obtain the compiled model.

Training model

To train the model use the "training_model.py" file. We found the best kernel sizes to be (1 x 4) and (4 x 1) for the subnetworks. To change the optimizer for the model, refer to the compile_model() method in the "carcinomaNetwork.py" file. Feel free to include/modify necessary callbacks and related training parameters.

Performance scores

To obtain the performance scores of the model one may use our "performance_viz.py" file, which includes methods to visualiza the confusion matrix. One may use keras-vis to visualize the grad-CAMs for the convolution layers. We also provide a method to plot the ROC curves for the model.

Requirements

Keras(v2.2.2)
tensorflow(v1.9.0)
keras-vis(v0.4.1)(optional-depends on user)
scikit-learn(v0.19.2)
mlxtend(v0.13.0)