ECE2024 Oral Communications Oral Communications 11: Adrenal and Cardiovascular Endocrinology | Part II (6 abstracts)
Prediction of adrenal masses nature through texture analysis and deep learning: Preliminary results from ENS@T RADIO-AI multicentric study
Lorenzo Tucci 1,2 , Giulio Vara 3,4 , Valentina Morelli 5 , Edelmiro Luis Menendez Torre 6 , Ulrich Dischinger 7 , Athina Markou 8 , Massimo Terzolo 9,10 , Ariadni Spyroglou 11 , Chiara Parazzoli 12 , Aresta Carmen 5 , Iacopo Chiodini 12,13 , Diego Rivas 6 , Alba Gutiérrez 6 , Wiebke Schlötelburg 14 , Krystallenia Alexandraki 11 , Soraya Puglisi 9,10 , Ilaria Improta 1,2 , Antonio De Leo 1,15 , Saverio Selva 16 , Laura Alberici 16 , Andrea De Giglio 1,17 , Maria Abbondanza Pantaleo 1,17 , Caterina Balacchi 18 , Cristina Mosconi 1,18 , Valentina Vicennati 1,2 , Uberto Pagotto 1,2 & Guido Di Dalmazi 1,2
1Alma Mater Studiorum-Bologna University, Medical and Surgical Sciences Department, Bologna, Italy; 2IRCCS S. Orsola-Malpighi Polyclinic, Endocrinology and Diabetes care and Prevention Unit, Bologna, Italy; 3Alma Mater Studiorum-Bologna University, Biomedical and Neuromotor Sciences Department, Bologna, Italy; 4Civil Hospital "Umberto I", Diagnostic and interventional radiology Unit, Lugo (RA), Italy; 5IRCCS Auxologico, Endocrinology and Metabolism Disease Unit, Milan, Italy; 6Asturias Central University Hospital, Endocrinology and Nutrition Service, Oviedo, Spain; 7University Hospital Würzburg, Department of Endocrinology and Diabetes, Würzburg, Germany; 8G. Gennimatas General Hospital, Endocrinology Department, Athens, Greece; 9Turin University, Biological and Clinical Sciences Department, Turin, Italy; 10San Luigi Gonzaga University Hospital, Endocrine Internal Medicine, Turin, Italy; 11Aretaieio Hospital, National and Kapodistrian University of Athens, 2nd Department of Surgery, Athens, Greece; 12University of Milan, Department of Biotechnology and Translational Medicine, Milan, Italy; 13Niguarda Cà Granda Hospital, Unit of Endocrinology, Milan, Italy; 14Würzburg University Hospital, Department of Nuclear Medicine, Würzburg, Germany; 15IRCCS S. Orsola-Malpighi Polyclinic, Anatomic Pathology Unit, Bologna, Italy; 16IRCCS S. Orsola-Malpighi Polyclinic, Division of Pancreatic and Endocrine Surgery, Bologna, Italy; 17IRCCS S. Orsola-Malpighi Polyclinic, Medical Oncology Unit, Bologna, Italy; 18IRCCS S. Orsola-Malpighi Polyclinic, Diagnostic and interventional abdominal-pelvic radiology Unit, Bologna, Italy
Background: Current parameters of conventional radiology have several limitations in defining the nature of adrenal masses. Radiomics, or texture analysis, has shown high diagnostic performance in recent pilot studies, although confirmatory studies are needed. Moreover, the effect of combination of radiomics with hormonal secretion on diagnostic performance is poorly explored.
Aim: To evaluate the accuracy of radiomics in predicting adrenal masses nature in a large cohort of patients.
Methods: In 7 European Centres, we retrospectively analyzed 794 adrenal masses from 2006 to 2023: 472 lipid-rich adenomas (LRA), 178 benign indeterminate adrenal masses (BIAM, with histological confirmation or radiologic stability at 6-12 months), 33 adrenocortical carcinomas (ACC), 45 pheochromocytomas, 48 adrenal metastases, and 18 adrenal cysts (confirmed by nuclear magnetic resonance or histology). Adrenal masses were also divided by hormonal secretion: 354 non secreting, 261 with mild autonomous cortisol secretion, 42 with Cushing’s syndrome, 45 pheochromocytomas and 94 with unknown secretion. ACC, pheochromocytomas and adrenal metastases were grouped together and labeled as malignant. Texture analysis was performed on unenhanced computerized tomography scan with LifeX software (version 7.2.0, ©LITO, France). We employed Deep Learning (DL) with 10-fold validation with hormonal secretion as factor and radiomic features as covariates.
Results: DL algorithms with radiomic parameters predicted the presence of malignant masses with average area under the receiver operating characteristic curve (AUROCC)=0.974, F1-score=0.801, sensitivity=92.7%, specificity=92.8%. Diagnostic accuracy was also high for subtyping adrenal masses: LRA with average AUROCC=0.979, F1-score=0.944, sensitivity=91.3%, specificity=96.8%, BIAM with average AUROCC=0.936, F1-score=0.746, sensitivity=93.1%, specificity=83.8%, ACC with average AUROCC=0.973, F1-score=0.405, sensitivity=100%, specificity=88%, pheochromocytoma with average AUROCC=0.908, F1-score=0.341, sensitivity=93.3%, specificity=78.6%, adrenal metastasis with average AUROCC=0.902, F1-score=0.367, sensitivity=97.9%, specificity=78.3%, and adrenal cysts with average AUROCC=0.899, F1-score=0.123, sensitivity=100%, specificity=67.1%. When hormonal secretion was added as a factor, the prediction of malignant adrenal masses by DL algorithms was even higher, with average AUROCC=0.999, F1-score 0.988, sensitivity 98.4%, specificity 99.9%. DL with radiomics and hormonal secretion predicted LRA with average AUROCC=0.983, F1-score=0.953, sensitivity=92.7%, specificity=97.3%, BIAM with average AUROCC=0.968, F1-score=0.816, sensitivity=96.3%, specificity=89.1%, ACC with average AUROCC=0.933, F1-score=0.318, sensitivity=96.4%, specificity=83.9%, pheochromocytoma with average AUROCC=0.941, F1-score=0.542, sensitivity=100%, specificity=89.2%, adrenal metastasis with average AUROCC=0.951, F1-score=0.599, sensitivity=97.9%, specificity=91.2%, and adrenal cysts with average AUROCC=0.848, F1-score=0.1, sensitivity=100%, specificity=63.4%.
Conclusion: Radiomic-based DL algorithm showed a high accuracy in predicting the presence of malignant adrenal masses and a heterogenous performance in predicting specific subtypes. The performance of the model was increased by adding the hormonal secretion to the models.
Article tools
My recent searches
My recently viewed abstracts
Authors
Endocrine Abstracts
ISSN 1470-3947 (print) | ISSN 1479-6848 (online)
© Bioscientifica 2024 |
Privacy policy |
Cookie settings
BiosciAbstracts
Bioscientifica Abstracts is the gateway to a series of products that provide a permanent, citable record of abstracts for biomedical and life science conferences.
Find out more