Túlio Guilherme Soares Marques
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Túlio Guilherme Soares Marques has a total of 18 co-authors. Here are the top co-authors by number of shared publications:
- Diana Rodrigues de Pina
São Paulo State University (UNESP)
5 shared publications - Raissa Alexia Camargo Guassu
Institute of Biosciences, São Paulo State University (UNESP), R. Prof. Dr. Antônio Celso Wagner Zanin, 250 - Distrito de Rubião Junior, Botucatu, SP, CEP: 18618-689, Brazil
4 shared publications - Daniel Molena Seraphim
Institute of Bioscience, São Paulo State University (UNESP), Botucatu, São Paulo, Brazil
2 shared publications - Jonas Nanni Menino
Institute of Bioscience, São Paulo State University (UNESP), Botucatu, São Paulo, Brazil
2 shared publications - Allan Felipe Fattori Alves
Institute of Biosciences, São Paulo State University (UNESP), Botucatu
1 shared publication - Ana Luiza Milani
Univ. Estadual Paulista Júlio de Mesquita Filho
1 shared publication
Latest Publications
- Open Access
- 4 June 2026
- Published by Wiley in Journal of Applied Clinical Medical Physics
- Vol. 27 (6) , e70650
Development and validation of a 3D printed phantom for image quality assessment in fluoroscopy
Background: Low‐contrast detectability is a key image‐quality parameter in fluoroscopy, directly influencing diagnostic accuracy and patient safety. Commercial phantoms such as the CDRAD provide reliable assessments but are costly and lack flexibility for routine use. Purpose: To develop and validate a low‐cost 3D‐printed phantom for fluoroscopy quality control, designed with low‐contrast performance similar to the commercial reference standard CDRAD, and to offer a reproducible framework for clinical implementation. This represents an original and accessible strategy for systematic fluoroscopy image‐quality assessment. Methods: The proposed 3D‐printed low‐contrast phantom (hereafter referred to as the LC phantom) was modeled in a Computer‐Aided Design (CAD) software and printed in polylactic acid (PLA) with 100% infill. It features a modular circular design adaptable to different fields of view. Quantitative validation included correlation between theoretical subject contrast (SC) and image‐based contrast‐to‐noise ratio (CNR), with comparisons to the CDRAD under clinically representative beam qualities (65–100 kVp), using additional copper filtration to produce hardened X‐ray spectra typical of fluoroscopic practice. Subjective comparison of both phantoms was performed through multicenter testing on four fluoroscopy systems (three mobile and one fixed). Results: CNR increased consistently with hole depth for both phantoms across all beam qualities, with linear regression models yielding coefficients of determination (R 2 ) ranging from 0.88 to 0.99. Visual evaluations yielded mean low‐contrast thresholds of 3.11 ± 0.26%, 2.66 ± 0.31%, and 2.70 ± 0.27% for 65, 80, and 100 kVp, respectively, with a maximum deviation of 0.10 SC units relative to the reference phantom, indicating reproducible and consistent performance across the evaluated beam qualities. Conclusions: The proposed LC phantom is a practical and affordable tool for routine fluoroscopy image quality assessment. Its methodology can be adapted to different beam qualities and institutional protocols, supporting harmonized and accessible image‐quality assessment in clinical practice.
- 4 June 2026
- Published by Wiley in Journal of Applied Clinical Medical Physics
- Vol. 27 (6) , e70650
Open AccessDevelopment and validation of a 3D printed phantom for image quality assessment in fluoroscopy
Background: Low‐contrast detectability is a key image‐quality parameter in fluoroscopy, directly influencing diagnostic accuracy and patient safety. Commercial phantoms such as the CDRAD provide reliable assessments but are costly and lack flexibility for routine use. Purpose: To develop and validate a low‐cost 3D‐printed phantom for fluoroscopy quality control, designed with low‐contrast performance similar to the commercial reference standard CDRAD, and to offer a reproducible framework for clinical implementation. This represents an original and accessible strategy for systematic fluoroscopy image‐quality assessment. Methods: The proposed 3D‐printed low‐contrast phantom (hereafter referred to as the LC phantom) was modeled in a Computer‐Aided Design (CAD) software and printed in polylactic acid (PLA) with 100% infill. It features a modular circular design adaptable to different fields of view. Quantitative validation included correlation between theoretical subject contrast (SC) and image‐based contrast‐to‐noise ratio (CNR), with comparisons to the CDRAD under clinically representative beam qualities (65–100 kVp), using additional copper filtration to produce hardened X‐ray spectra typical of fluoroscopic practice. Subjective comparison of both phantoms was performed through multicenter testing on four fluoroscopy systems (three mobile and one fixed). Results: CNR increased consistently with hole depth for both phantoms across all beam qualities, with linear regression models yielding coefficients of determination (R 2 ) ranging from 0.88 to 0.99. Visual evaluations yielded mean low‐contrast thresholds of 3.11 ± 0.26%, 2.66 ± 0.31%, and 2.70 ± 0.27% for 65, 80, and 100 kVp, respectively, with a maximum deviation of 0.10 SC units relative to the reference phantom, indicating reproducible and consistent performance across the evaluated beam qualities. Conclusions: The proposed LC phantom is a practical and affordable tool for routine fluoroscopy image quality assessment. Its methodology can be adapted to different beam qualities and institutional protocols, supporting harmonized and accessible image‐quality assessment in clinical practice.
- Open Access
- 1 June 2026
- Published by Revista Brasileira de Fisica Medica in Revista Brasileira de Física Médica
- Vol. 20
Comparison between Semi-Automated and Manual Methodologies for MTF Determination in a Digital Radiography System
Spatial resolution, although commonly assessed qualitatively, can be quantified by determining the system's Modulation Transfer Function (MTF). The literature is extensive regarding methods for determining MTF in digital radiology equipment. Among the methodologies, there are semi-automatic methods, which require image analysis by dedicated software, and manual methods, which require the operator to position regions of interest (ROIs). In this study, three MTF measurement methods discussed in the literature – two semi-automatic and one manual – were compared using the RQA 5 beam quality of a Konica Altus DR digital radiology system in a large hospital. The semi-automatic methods showed strong agreement with each other, with variations between 0.01 and 5.49 MTF units (%), while the manual method showed greater divergence from the other two, varying between 0.04 and 6.14 MTF units (%), depending on the fine or broad focus. It is concluded that semi-automatic methods are advantageous in terms of reproducibility and efficiency when compared to the manual method. It is the medical physicist's responsibility to weigh the applicability of each methodology within the reality of their institution.
- 1 June 2026
- Published by Revista Brasileira de Fisica Medica in Revista Brasileira de Física Médica
- Vol. 20
Open AccessComparison between Semi-Automated and Manual Methodologies for MTF Determination in a Digital Radiography System
Spatial resolution, although commonly assessed qualitatively, can be quantified by determining the system's Modulation Transfer Function (MTF). The literature is extensive regarding methods for determining MTF in digital radiology equipment. Among the methodologies, there are semi-automatic methods, which require image analysis by dedicated software, and manual methods, which require the operator to position regions of interest (ROIs). In this study, three MTF measurement methods discussed in the literature – two semi-automatic and one manual – were compared using the RQA 5 beam quality of a Konica Altus DR digital radiology system in a large hospital. The semi-automatic methods showed strong agreement with each other, with variations between 0.01 and 5.49 MTF units (%), while the manual method showed greater divergence from the other two, varying between 0.04 and 6.14 MTF units (%), depending on the fine or broad focus. It is concluded that semi-automatic methods are advantageous in terms of reproducibility and efficiency when compared to the manual method. It is the medical physicist's responsibility to weigh the applicability of each methodology within the reality of their institution.
- Open Access
- 7 May 2026
- Published by Revista Brasileira de Fisica Medica in Revista Brasileira de Física Médica
- Vol. 20
Deep Learning-Based Segmentation of Hepatic Structures: A Systematic Review of Model Architectures and Clinical Applications
Liver cancer is a major public health problem and remains among the most incident neoplasms worldwide, with approximately 866,000 new cases reported in 2022, of which hepatocellular carcinoma accounts for the majority of diagnoses. Anatomical characterization of the liver and its lesions relies largely on computed tomography (CT) and magnetic resonance imaging (MRI), whose segmentation is still predominantly manual or semi-automatic. This process is time-consuming and may lead to variability between different observers or even for the same professional at different time points. This limitation has motivated the adoption of deep learning (DL)–based methods, particularly to improve the accuracy and reproducibility of hepatic segmentation. This review analyzed 16 studies selected based on predefined keywords and inclusion criteria, aiming to identify methodological trends, predominant network architectures, and reported clinical applications. A strong predominance of U-Net–based models and their variants was observed (9/16), alongside an increasing use of three-dimensional approaches (11/16) and a focus on vascular segmentation (6/16) or combined liver-and-vessel segmentation (6/16). Methodological strategies were equally divided between combined approaches (8/16) and single-step methods (8/16). Despite the high technical performance reported by several authors, the studies showed substantial heterogeneity in acquisition protocols, evaluation metrics, and dataset size, which hampers direct comparison and limits the translation of these results into clinical practice. It can be concluded that, although deep learning represents a significant advance in hepatic segmentation, methodological standardization and external model validation are still required for its consolidation in the clinical setting.
- 7 May 2026
- Published by Revista Brasileira de Fisica Medica in Revista Brasileira de Física Médica
- Vol. 20
Open AccessDeep Learning-Based Segmentation of Hepatic Structures: A Systematic Review of Model Architectures and Clinical Applications
Liver cancer is a major public health problem and remains among the most incident neoplasms worldwide, with approximately 866,000 new cases reported in 2022, of which hepatocellular carcinoma accounts for the majority of diagnoses. Anatomical characterization of the liver and its lesions relies largely on computed tomography (CT) and magnetic resonance imaging (MRI), whose segmentation is still predominantly manual or semi-automatic. This process is time-consuming and may lead to variability between different observers or even for the same professional at different time points. This limitation has motivated the adoption of deep learning (DL)–based methods, particularly to improve the accuracy and reproducibility of hepatic segmentation. This review analyzed 16 studies selected based on predefined keywords and inclusion criteria, aiming to identify methodological trends, predominant network architectures, and reported clinical applications. A strong predominance of U-Net–based models and their variants was observed (9/16), alongside an increasing use of three-dimensional approaches (11/16) and a focus on vascular segmentation (6/16) or combined liver-and-vessel segmentation (6/16). Methodological strategies were equally divided between combined approaches (8/16) and single-step methods (8/16). Despite the high technical performance reported by several authors, the studies showed substantial heterogeneity in acquisition protocols, evaluation metrics, and dataset size, which hampers direct comparison and limits the translation of these results into clinical practice. It can be concluded that, although deep learning represents a significant advance in hepatic segmentation, methodological standardization and external model validation are still required for its consolidation in the clinical setting.
- Open Access
- 1 January 2026
- Published by XMLink in Investigative Magnetic Resonance Imaging
- Vol. 30 (3) , 178-184
Radiomics-Based Prediction of Pituitary Macroadenoma Consistency Using T2-Weighted Magnetic Resonance Imaging With Heterogeneous Clinical Data
Marques TG, et al. Investig Magn Reson Imaging. 2026 Sep;30(3):178-184. https://doi.org/10.13104/imri.2026.0014
- 1 January 2026
- Published by XMLink in Investigative Magnetic Resonance Imaging
- Vol. 30 (3) , 178-184
Open AccessRadiomics-Based Prediction of Pituitary Macroadenoma Consistency Using T2-Weighted Magnetic Resonance Imaging With Heterogeneous Clinical Data
Marques TG, et al. Investig Magn Reson Imaging. 2026 Sep;30(3):178-184. https://doi.org/10.13104/imri.2026.0014
- Open Access
- 30 December 2022
- Published by Revista Brasileira de Fisica Medica in Revista Brasileira de Física Médica
- Vol. 16, 678
Dose Efetiva e Nível de Referência de Dose (DRL) em radiologia em um hospital terciário
Uma das maneiras de promover a otimização em proteção radiológica é implementando o conceito de Nível de Referência de Dose (DRL). Diversas são as formas de estimar a dose de radiação, dentre elas a estimativa do produto área-dose (DAP). Neste estudo, foi estimada a dose efetiva em exames radiológicos com o objetivo de determinar o DRL local. Para isso, os dados de DAP foram coletados e filtrados para o modelo de paciente padrão. Todos os procedimentos avaliados foram realizados em equipamentos de radiologia convencional com medidor de DAP embutido e devidamente calibrado, dentre eles: tórax anteroposterior (AP), tórax lateral (LAT) e tórax póstero-anterior (PA), abdômen AP, pelve AP e coluna lombar AP e LAT. O software PCXMC 2.0 foi empregado na simulação computacional de exames radiográficos para simular a dose efetiva no paciente para as incidências utilizadas. No total, após verificação dos dados, foram selecionados 167 valores de DAP e o número de incidências realizadas variou de 7 a 51 para diferentes projeções. A maior quantidade de incidências foi coletada para tórax LAT e a menor quantidade para coluna lombar LAT. Foram analisados 1º quartil, mediana e 3º quartil, bem como DAP mínimo e máximo e o valor interquartil (IQV) para cada conjunto de dados. As simulações computacionais resultaram em valores de dose efetiva compatíveis com a literatura. Por se tratar de um trabalho inicial, alguns pontos devem ser levados em consideração, como: o serviço faz parte de um hospital escola com profissionais em formação; existem cartas de técnicas disponíveis, porém os equipamentos digitais foram recém instalados; o estudo deve ser aprofundado, com número maior de incidências, para que os protocolos sejam bem estabelecidos no serviço. Este estudo serve de referência para outros serviços que desejam realizar o levantamento do nível de referência de diagnóstico para protocolos de radiologia convencional.
- 30 December 2022
- Published by Revista Brasileira de Fisica Medica in Revista Brasileira de Física Médica
- Vol. 16, 678
Open AccessDose Efetiva e Nível de Referência de Dose (DRL) em radiologia em um hospital terciário
Uma das maneiras de promover a otimização em proteção radiológica é implementando o conceito de Nível de Referência de Dose (DRL). Diversas são as formas de estimar a dose de radiação, dentre elas a estimativa do produto área-dose (DAP). Neste estudo, foi estimada a dose efetiva em exames radiológicos com o objetivo de determinar o DRL local. Para isso, os dados de DAP foram coletados e filtrados para o modelo de paciente padrão. Todos os procedimentos avaliados foram realizados em equipamentos de radiologia convencional com medidor de DAP embutido e devidamente calibrado, dentre eles: tórax anteroposterior (AP), tórax lateral (LAT) e tórax póstero-anterior (PA), abdômen AP, pelve AP e coluna lombar AP e LAT. O software PCXMC 2.0 foi empregado na simulação computacional de exames radiográficos para simular a dose efetiva no paciente para as incidências utilizadas. No total, após verificação dos dados, foram selecionados 167 valores de DAP e o número de incidências realizadas variou de 7 a 51 para diferentes projeções. A maior quantidade de incidências foi coletada para tórax LAT e a menor quantidade para coluna lombar LAT. Foram analisados 1º quartil, mediana e 3º quartil, bem como DAP mínimo e máximo e o valor interquartil (IQV) para cada conjunto de dados. As simulações computacionais resultaram em valores de dose efetiva compatíveis com a literatura. Por se tratar de um trabalho inicial, alguns pontos devem ser levados em consideração, como: o serviço faz parte de um hospital escola com profissionais em formação; existem cartas de técnicas disponíveis, porém os equipamentos digitais foram recém instalados; o estudo deve ser aprofundado, com número maior de incidências, para que os protocolos sejam bem estabelecidos no serviço. Este estudo serve de referência para outros serviços que desejam realizar o levantamento do nível de referência de diagnóstico para protocolos de radiologia convencional.