Home > Journals > Journal of Radiological Review > Past Issues > Il Giornale Italiano di Radiologia Medica 2018 Luglio-Agosto;5(4) > Il Giornale Italiano di Radiologia Medica 2018 Luglio-Agosto;5(4):488-95

CURRENT ISSUE
 

JOURNAL TOOLS

eTOC
To subscribe
Recommend to your librarian
 

ARTICLE TOOLS

Publication history
Reprints
Permissions
Cite this article as

 

ORIGINAL ARTICLE   

Il Giornale Italiano di Radiologia Medica 2018 Luglio-Agosto;5(4):488-95

DOI: 10.23736/S2283-8376.18.00108-0

Copyright © 2018 EDIZIONI MINERVA MEDICA

language: Italian

Neural networks in Crohn disease patterns classification: the role of kernel support vector machine

Maria Chiara TERRANOVA , Federica LI POMI, Maria Teresa ANZELMO, Alessia VINCI, Agata CRAPANZANO, Alberto CALANDRA, Sergio SALERNO, Giuseppe LO RE, Salvatore VITABILE, Massimo MIDIRI

Department of Medical Biopathology and Biotechnologies, University of Palermo, Palermo, Italy


PDF


BACKGROUND: We propose a new technique for pattern classification in patients affected by Crohn’s disease (CD). The proposed technique is based on a kernel support vector machine (KSVM) and it adopts a stratified K-fold cross-validation strategy to enhance the KSVM classifier reliability. Traditional manual classification methods require radiological expertise and they usually are very time-consuming.
METHODS: Dataset composed of 300 patients has been extracted by three expert radiologists for KSVM training and validation. Each patient was codified by 22 extracted qualitative features and classified as positive or negative of different patterns as the related histological specimen results. The effectiveness of the proposed technique has been proved using a real human patient dataset collected at the University of Palermo Polyclinic Hospital (UPPH) and composed of 300 patients.
RESULTS: The achieved results (sensitivity: 94.80%; specificity: 100.00%; negative predictive value: 95.06%; precision: 100.00%; accuracy: 97.40%; error: 2.60%) show that the proposed technique results are better than the manual reference methods reported in literature (sensitivity: 93.00%; specificity: 90.00%) with a greater usability and degree of acceptance.
CONCLUSIONS: This KSVM-based pattern classification technique has the advantage of including a validation phase. Moreover, its cross-validation strategy increases its generalization ability without over-fitting nor sensitivity to the selected test dataset. Lastly, this tool’s acceptance rate in clinical practice is considerably high.


KEY WORDS: Crohn disease - X-ray computed tomography - Neural pathways - Machine learning

top of page