Deep learning system for the automatic classification of normal and dysplastic peripheral blood cells as a suport tool for the diagnosis

• Autores: Andrea Milena Acevedo Lipes
• Directores de la Tesis: Anna Merino (dir. tes.) , José Julián Rodellar Benedé (codir. tes.)
• Lectura: En la Universitat de Barcelona ( España ) en 2021
• Idioma: español
• Tribunal Calificador de la Tesis: Raúl Benítez Iglesias (presid.) , Manuel Morales Ruiz (secret.) , Maite Serrano Querol (voc.)
• Texto completo no disponible (Saber más ...)
• Resumen

  • Clinical pathologists identify visually many morphological features to characterize the different normal cells, as well as the abnormal cell types whose presence in peripheral blood is the evidence of serious diseases. Disadvantages of visual morphological analysis are that it is time consuming, needs expertise to perform an objective review of the smears and is prone to inter-observer variability. Also, most of the morphological descriptions are given in qualitative terms and there is a lack of quantitative measures. The general objective of this thesis is the automatic recognition of normal and dysplastic cells circulating in blood in myelodysplastic syndromes using convolutional neural networks and digital image processing techniques. In order to accomplish this objective, this work starts with the design and development of a mysql database to store information and images from patients and the development of a first classifier of four groups of cells, using convolutional neural networks as feature extractors. Then, a high-quality dataset of around 17,000 images of normal blood cells is compiled and used for the development of a recognition system of eight groups of blood cells. In this work, we compare two transfer learning approaches to find the best to classify the different cell types. In the second part of the thesis, a new convolutional neural network model for the diagnosis of myelodysplastic syndromes is developed. This model was validated by means of a proof of concept. It is considered among the first models that have been built for diagnosis support. The final work of the thesis is the integration of two convolutional networks in a modular system for the automatic classification of normal and abnormal cells. The methodology and models developed constitute a step forward to the implementation of a modular system to recognize automatically all cell types in a real setup in the laboratory. Keywords: Blood morphology; Convolutional neural networks; Blood cells automatic recognition; Myelodysplastic syndromes; Diagnosis support.