Eduardo Pérez – Knowledge Discovery and Intelligent Systems – KDIS

AUTOMATIC DIAGNOSIS OF MELANOMA WITH MODERN MACHINE LEARNING TECHNIQUES.

BASIC INFORMATION

Ph.D. Student: Eduardo Pérez
Advisors: Sebastián Ventura
Defended on: December 2022
Keywords: early diagnosis of melanoma, deep learning models, flexible data representations
Digital version

THESIS PROPOSAL

In the last two decades, the application of data mining and machine learning techniques for automating medical diagnosis has gained increasing attention by the scientific community. The main reason for applying this type of computational methods lies in their ability to extract useful knowledge in scenarios where it turns difficult, or even impossible, to draw a conclusion by. Furthermore, the increase of the overall patient information stored, for example in electronic health records (EHR), as well as the amount of information generated by most of the new diagnostic tests (gene arrays, MRI, etc.) is motivating a significant growth of the application of data analysis techniques, including big data, as a support tool for the analysis and automatic diagnosis in biomedicine.

The main goal of this thesis is the development of modern machine learning methods for the automatic (or semi-automatic) diagnosis of melanoma at early stages. This type of skin cancer has an increasing incidence in white people, causing close to 90% of skin cancer mortality. Furthermore, the incidence rate in Europe is around 10-25 new melanoma cases per 100,000 inhabitants; in USA, it is about 20-30 per 100,000 inhabitants; whereas in Australia it affects over 50-60 per 100,000 inhabitants.

The main contributions of this Ph.D. thesis are summarized in the following points:

An extensive experimental study about melanoma diagnosis was carried out, which compared several baseline CNNs to the same ones but evaluating different methods and techniques: a) optimization algorithms, b) weight balancing, c) transfer learning, d) data augmentation. Overall, a consensus was not found when analyzing the optimization algorithms and weight balancing techniques – specific behaviors were found in each CNN model. On the other hand, transfer learning and data augmentation proved to be suitable techniques for achieving significantly better diagnostic performance.
A genetic algorithm for the automatic selection and training process of CNN models is proposed. The main aim is to select the CNNs that best contribute to the ensemble, rather than the individual level. In this work a genetic algorithm is used to find the best set of CNNs to build an ensemble-based architecture.
A first approach about a multimodal predictive model is proposed. The method combines clinical and imaging data, genetic algorithm, and classical techniques such as data augmentation and transfer learning. In addition, several fusion strategies are analyzed, as well as a search regarding how to combine features and predictions. Experimental results show better performance than an imaging-based CNN model.
We aimed to design a simpler architecture, which is composed of a single predictive block and uses a recently proposed approach to detect spatial hierarchies between entities within an image. In addition, a convolution-based computational block is initially used in order to extract abstract features before passing them to the main predictive block. The main block is based on CAPSNET, whose equivariant characteristics make it more attractive since the model is capable of detecting the rotation or proportion change and adapt itself in a way that the objects are internally represented as vectors.
In order to overcome the limited training data, a Progressive Growing Generative Adversarial Network architecture is proposed. The proposal has been evaluated qualitatively and quantitatively through the use of an extensive experimental study on sixteen dermoscopic and non-dermoscopic skin image datasets, illustrating its effectiveness in the diagnosis of melanoma.
We propose a pipeline to train CNN models by analyzing how informative each sample is. We hypothesize that a better performance in least number of epochs could be achieved if CNN models are able to analyze beforehand from where training is performed. A custom active learning approach guides the training process, where the convolutional architecture is benefited from its uncertainty about individual skin images.

FUNDS

The development of this thesis is being supported by:

Strategic Action in Health 2017 of Spain, i-PFIS contract – IFI17/00015.
Spanish Ministry of Science and Competitiveness, project TIN2017-83445-P.
Spanish Ministry of Science and Innovation and the European Regional Development Fund, under project PID2020-115832GB-I00

PUBLICATIONS ASSOCIATED WITH THIS THESIS

INTERNATIONAL JOURNALS

Pérez, E., & Ventura, S. (2023). Progressive growing of Generative Adversarial Networks for improving data augmentation and skin cancer diagnosis. Artificial Intelligence in Medicine, 141, 102556.
Pérez, E., & Ventura, S. (2022). A framework to build accurate Convolutional Neural Network models for melanoma diagnosis. Knowledge-Based Systems, 110157.
Pérez, E., & Ventura, S. (2021). An ensemble-based convolutional neural network model powered by a genetic algorithm for melanoma diagnosis. Neural Computing and Applications.
Pérez, E., & Ventura, S. (2021). Melanoma Recognition by Fusing Convolutional Blocks and Dynamic Routing between Capsules. Cancers, 13(19).
Pérez, E., Reyes, O., & Ventura, S. (2021). Convolutional neural networks for the automatic diagnosis of melanoma: An extensive experimental study. Medical Image Analysis, 67.

INTERNATIONAL CONFERENCES

Perez, E., & Ventura, S. (2022). Multi-view Deep Neural Networks for multiclass skin lesion diagnosis. 2022 IEEE International Conference on Omni-Layer Intelligent Systems (COINS), 1–6.
Reyes, O., Pérez, E., & Ventura, S. (2019). Performing melanoma diagnosis by an effective convolutional architecture. The International Skin Imaging Collaboration (ISIC) Challenge, ISIC 2019, 4.