ETD

Digital archive of theses discussed at the University of Pisa

 

Thesis etd-10202020-094134


Thesis type
Tesi di specializzazione (4 anni)
Author
CASTELLANA, ROBERTO
URN
etd-10202020-094134
Thesis title
A NOVEL DEEP LEARNING METHOD BASED ON CONVOLUTIONAL NEURAL NETWORK FOR INTERSTITIAL LUNG DISEASE HRCT QUANTIFICATION
Department
RICERCA TRASLAZIONALE E DELLE NUOVE TECNOLOGIE IN MEDICINA E CHIRURGIA
Course of study
RADIODIAGNOSTICA
Supervisors
relatore Prof. Caramella, Davide
Keywords
  • interstitial lung disease
  • ILD
  • idiopathic pulmonary fibrosis
  • HRCT
  • convolutional neural network
  • quantification
Graduation session start date
07/11/2020
Availability
Full
Summary
Interest for deep learning in radiology has increased tremendously in the past decade due to the high achievable performance for various computer vision tasks such as detection, segmentation, classification, monitoring, and prediction.
This was a retrospective study with the aim of assessing the proficiency of a deep learning method based on a novel convolutional neural network, named UIP-net, for quantifying the fibrotic extent on chest HRCTs of patients with idiopathic pulmonary fibrosis (IPF). Chest HRCT volumes with interstitial lung disease (ILD) measured by UIP-net were compared with those calculated with CALIPER, a well-known texture analysis software for ILD. Furthermore, the relations between the fibrotic extent assessed by both methods and the pulmonary function tests (PFTs), i.e. the forced vital capacity (FVC), the diffusing capacity of the lung for carbon monoxide (DLCO) and the forced expiratory volume in 1 second (FEV1) were evaluated.
Sixty patients with HRCT and PFTs were included in the study. HRCTs were first analysed using UIP-net and then with CALIPER in order to calculate the pulmonary volumes with ILD.
The intraclass correlation coefficient between UIP-net and CALIPER measurements was assessed.
Spearman’s correlation analysis of the covariates was used to evaluate a correlation between volumes with ILD measured by both methods and the PFTs.
An excellent intraclass correlation coefficient of 0.973 between UIP-net and CALIPER was found, with a 95% confidence interval from 0.768 to 0.992.
Furthermore, negative significant correlations were evidenced between volumes with ILD measured by CALIPER and FVC (Rho = -0.403; p value < 0.001), DLCO (Rho = -0.476; p value = 0.009), and FEV1 (Rho = -0.322; p value = 0.003). The volumes with ILD measured by UIP-net were correlated with FVC (Rho = -0.415; p value = 0.007) and DLCO (Rho = -0.434; p value = 0.007).
In conclusion, UIP-net is a reliable deep-learning method for quantifying the pulmonary fibrosis on chest HRCTs and the measured volumes with ILD correlate with FVC and DLCO.
File