| Contributors: |
Parmeggiani, Domenico; Napolitano, Vincenzo; Moccia, Giancarlo; d’Orlando, Vittorio; Sperlongano, Pasquale; Luongo, Pasquale; Miele, Francesco; Allaria, Alfredo; Torelli, Francesco; Marrone, Stefano; Gravina, Michela; Sansone, Carlo; Bollino, Ruggiero; Bassi, Paola; Sciarra, Antonella; Francesco, Lucido; Di Domenico, Marina; Della Monica, Paola; Colapietra, Federica; Ballini, Andrea; Docimo, Ludovico; Agresti, Massimo; Parmeggiani, D.; Napolitano, V.; Moccia, G.; D'Orlando, V.; Sperlongano, P.; Luongo, P.; Miele, F.; Allaria, A.; Torelli, F.; Marrone, S.; Gravina, M.; Sansone, C.; Bollino, R.; Bassi, P.; Sciarra, A.; Francesco, L.; Di Domenico, M.; Monica, P. D.; Colapietra, F.; Ballini, A.; Docimo, L.; Agresti, M. |
| Description: |
The use of methods based on machine learning algorithms and in partic-ular the use of Neural Networks first and Convolutional Neural Networks (CNNs) then have been trying for years to contribute to the imaging detection process in the medical field, and one of the most captivating challenges is certainly the possibility of contributing to the early diagnosis of colorectal cancer pathologies. The sample under examination collected consists of 106 patients (57F and 42 M), and all subjected to pancolonoscopy and endoscopic biopsy with anatomopathological findings of 54 low-grade dysplasias, 26 high-grade dysplasias, and 26 hyperplastic cancers. The Synergy-Net platform involves the use of two modules, respectively, for the detec-tion of any potentially neoplastic lesion and for its classification as benign/malignant: the first module is a Single-Shot Detector-type Neural Network, which has the task of analyze the video frame-by-frame in search of potentially neoplastic tissues, and the second module had the aim of classifying suspicious tissues in different patholo-gies. In more detail, AlexNet was used to distinguish the lesion between different classes. The network involved was pre-trained with the ImageNet dataset, and the task was solved with fine tuning. Metric precision (MAP) was used as the metric eval-uation. These methods allowed performance using matched pair of results, giving a set of pairs for several cross-validation. The Synergy-Net digital platform, in partic-ular the dedicated Workstation, made available to the expert endoscopist operator, was able to identify: Low-Grade Dysplasia 40/54 (74.07% with 0 false positives), High-Grade Dysplasia Grade 26/26 (100%), Hyperplastic Cancers 22/26 (84.61% with only 2 false negatives), 7.69%: the results of the injury detection module. The diagnosis is carried out through a system that uses a module to identify the lesions and by a classifier that carries out the diagnosis. The use of Synergy-Net proposes to automatically recognize polyps during a colonoscopic ... |