Clinical validation of the M5L lung computer-assisted detection system

Background: Early diagnosis of lung cancer could be crucial in trying to reduce the mortality. Alongside screening programmes, several Computer-Assisted Detection (CAD) systems for the automatic detection of pulmonary nodules, were developed, in order to support radiologists in the diagnosis. Methods: The M5L CAD, developed by the INFN in collaboration with CEADEN (Habana, Cuba) combines the results of two independent algorithms and provides a framework for further extension to others. The sensitivity is about 80% in the 4–6 false positive findings/scan range, which, considering the fact M5L was applied in a clinical-like approach, with no optimization and no data selection, is satisfactory. The development team tackled the issue of making it available to the largest possible user community. Therefore, a Web/Cloud prototype was designed and implemented: CT scans can be uploaded asynchronously by ICT staff in health facilities, while the M5L results are directly sent to the radiologist e-mail accounts in DICOM-compatible format. Clinical validation on oncological patients undergoing staging or restaging has recently started at IRCCS Candiolo, Italy. A panel formed by three radiologists with different level of expertise independently annotates the cases through the M5L web interface in first-reader mode. Each annotation includes not only spatial information about the nodule, but also several information about its features (e.g. malignancy, speculation). After each annotation is completed, M5L results are prompted to the radiologist, who reviews the first-reading results. Results: Preliminary results are showing a CAD sensitivity greater than 80% at about 4 FP/ scan which is found to be higher than the one obtained within the LIDC data-set. Furthermore, the usage of CAD is sensitively increasing the performance of the radiologists. Conclusions: We are proposing a CAD framework in clinical routine which frees the user from buying additional software / hardware to access CAD results. Furthermore, this approach not only allows to share CAD results and annotations between radiologists not necessarily belonging to the same institution, but also allows to combine different CADs developed by different groups with no particular effort. Legal entity responsible for the study: INFN Turin Funding: Ministero della Istruzione e Ricerca Disclosure: All authors have declared no conflicts of Clinical validation on oncological patients undergoing staging or restaging has recently started at IRCCS Candiolo, Italy. A panel formed by three radiologists with different level of expertise independently annotates the cases through the M5L web interface in first-reader mode. Each annotation includes not only spatial information about the nodule, but also several information about its features (e.g. malignancy, speculation). After each annotation is completed, M5L results are prompted to the radiologist, who reviews the first-reading results. Results: Preliminary results are showing a CAD sensitivity greater than 80% at about 4 FP/ scan which is found to be higher than the one obtained within the LIDC data-set. Furthermore, the usage of CAD is sensitively increasing the performance of the radiologists. Conclusions: We are proposing a CAD framework in clinical routine which frees the user from buying additional software / hardware to access CAD results. Furthermore, this approach not only allows to share CAD results and annotations between radiologists not necessarily belonging to the same institution, but also allows to combine different CADs developed by different groups with no particular effort.