Scattering networks: efficient 2D implementation and application to melanoma classification

Machine learning is an approach to solving complex tasks. Its adoption is growing steadily and the several research works active on the field are publishing new interesting results regularly. In this work, the scattering network representation is used to transform raw images in a set of features convenient to be used in an image classification task, a fundamental machine learning application. This representation is invariant to translations and stable to small deformations. Moreover, it does not need any sort of training, since its parameters are fixed and only some hyper-parameters must be defined. A novel, efficient code implementation is proposed in this thesis. It leverages on the power of GPUs parallel architecture in order to achieve performance up to 20× faster than earlier codes, enabling near real-time applications. The source code of the implementation is also released open-source. The scattering network is then applied on a complex dataset of textures to test the behaviour in a general classification task. Given the conceptual complexity of the database, this unspecialized model scores a mere 32.9 % of accuracy. Finally, the scattering network is applied to a classification task of the medical field. A dataset of images of skin lesions is used in order to train a model able to classify malignant melanoma against benign lesions. Malignant melanoma is one of the most dangerous skin tumor, but if discovered in early stage there are generous probabilities to recover. The trained model has been tested and an interesting accuracy of 70.5 % (sensitivity 72.2 %, specificity 70.0 %) has been reached. While not being values high enough to permit the use of the model in a real application, this result demonstrates the great capabilities of the scattering network representation.