Carregant...
Fitxers
Tipus de document
TesiVersió
Versió publicadaData de publicació
Llicència de publicació
Si us plau utilitzeu sempre aquest identificador per citar o enllaçar aquest document: https://hdl.handle.net/2445/205984
Optimization of neural networks for deep learning and applications to CT image segmentation
Títol de la revista
Autors
Director/Tutor
ISSN de la revista
Títol del volum
Resum
[eng] During the last few years, AI development in deep learning has been going so fast that even important researchers, politicians, and entrepreneurs are signing petitions to try to slow it down. The newest methods for natural language processing and image generation are achieving results so unbelievable that people are seriously starting to think they can be dangerous for society. In reality, they are not dangerous (at the moment) even if we have to admit we reached a point where we have no more control over the flux of data inside the deep networks. It is impossible to open a modern
deep neural network and interpret how it processes the information and, in many cases, explain how or why it gives back that particular result. One of the goals of this doctoral work has been to study the behavior of weights in convolutional neural networks and in transformers. We hereby present a work that demonstrates how to invert 3x3 convolutions after training a neural network able to learn how to classify images, with the future aim of having precisely invertible convolutional neural networks. We demonstrate that a simple network can learn to classify images on an open-source dataset without loss in accuracy, with respect to a non-invertible one. All that with the ability to reconstruct the original image without detectable error
(on 8-bit images) in up to 20 convolutions stacked in a row. We present a thorough comparison between our method and the standard. We tested the
performances of the five most used transformers for image classification on an open- source dataset. Studying the embedded matrices, we have been
able to provide two criteria that can help transformers learn with a training time reduction of up to 30% and with no impact on classification accuracy.
The evolution of deep learning techniques is also touching the field of digital health. With tens of thousands of new start-ups and more than 1B $ of investments only in the last year, this field is growing rapidly and promising to revolutionize healthcare. In this thesis, we present several neural networks for the segmentation of lungs, lung nodules, and areas affected by pneumonia induced by COVID-19, in chest CT scans. The architecturesm we used are all residual convolutional neural networks inspired by UNet and Inception. We customized them with novel loss functions and layers
studied to achieve high performances on these particular applications. The errors on the surface of nodule segmentation masks are not over 1mm in more than 99% of the cases. Our algorithm for COVID-19 lesion detection has a specificity of 100% and overall accuracy of 97.1%. In general, it surpasses the state-of-the-art in all the considered statistics, using UNet as a benchmark. Combining these with other algorithms able to detect and predict lung cancer, the whole work was presented in a European innovation program and judged of high interest by worldwide experts.
With this work, we set the basis for the future development of better AI tools in healthcare and scientific investigation into the fundamentals of deep learning.
Descripció
Matèries (anglès)
Citació
Col·leccions
Citació
PEZZANO, Giuseppe. Optimization of neural networks for deep learning and applications to CT image segmentation. [consulta: 6 de desembre de 2025]. [Disponible a: https://hdl.handle.net/2445/205984]