Faculty of Medicine
Library of the School of Health Sciences

2024-04-11

2024

Kalatzis Dimitrios

Ευστάθιος Ευσταθόπουλος, Καθηγητής, Ιατρική Σχολή, ΕΚΠΑ
Αικατερίνη Μαλαγάρη, Καθηγήτρια, Ιατρική Σχολή, ΕΚΠΑ
Νικόλαος Κελέκης, Καθηγητής, Ιατρική Σχολή, ΕΚΠΑ
Παντελής Καραΐσκος, Καθηγητής, Ιατρική Σχολή, ΕΚΠΑ
Παναγιώτης Μπαμίδης, Καθηγητής, Ιατρική Σχολή, ΑΠΘ
Καλλιόπη Πλατώνη, Αναπληρώτρια Καθηγήτρια, Ιατρική Σχολή, ΕΚΠΑ
Κωνσταντίνος Λουκάς, Αναπληρωτής Καθηγητής, Ιατρική Σχολή, ΕΚΠΑ

Μέθοδοι Βαθιάς Μάθησης (Deep Learning) για την ανάλυση φασματοσκοπικών μεθόδων

Greek

Deep Learning methods for the analysis of spectroscopic methods

Raman spectroscopy (RS) has emerged as a powerful tool in the medical field for biochemical analysis and tissue discrimination. The integration of artificial intelligence (AI) algorithms with RS has significantly enhanced its capabilities, enabling precise and real-time spectral data analysis.
In this PhD dissertation, artificial intelligence algorithms were applied to distinguish spectral fingerprints of human health/cancerous tissues of the colon and rectum. The ultimate goal is the personalized rapid diagnosis in open surgery of cancer structures /cells in real-time.
Initially, spectra specimens from 22 patients undergoing open colorectal surgery, both from healthy and cancerous tissues, were collected to create the spectral library and then to discriminate the spectral fingerprints using AI.
Our research involved proposing pre-processing methods and algorithms to enhance classification outcomes, which included techniques like baseline correction, L2 normalization, filtering, and PCA. These enhancements resulted in an impressive overall accuracy improvement of 16.1%.
Also, a thorough ablation study to compare machine learning and deep learning algorithms focuses on advancing the clinical applicability of RS. Machine learning models proved effective in classifying both normal and abnormal tissues. Deep learning models, particularly the 1D-CNN model, performed better in classifying abnormal cases. In addition to these advancements, we addressed the challenge of large data requirements for deep learning methods by developing transfer learning models.
The pre-trained models were successfully trained on a Raman open database with pathogen bacteria spectra, achieving an 88% accuracy in discriminating healthy and cancerous tissues.
Also leverage the capability of a 1D-CNN with Transfer Learning, to classify in-vivo Raman spectra from mice achieving a remarkable 91,2% accuracy. The models overcome the limitations of the large data collection and demonstrate their effectiveness in ex-vivo and in-vivo settings. Overall, all results brought RS one step closer to clinical application as an auxiliary tool for real-time biopsy and surgical guidance.

Health Sciences

Raman spectroscopy, Colorectal cancer, Tissue discrimination, Machine learning, Deep learning, Transfer learning

No

0

Yes

102

117

File access is restricted only to the intranet of UoA.

https://pergamos.lib.uoa.gr/uoa/dl/object/3396108