Relationship between object detection and image classification stages in pap test

Autores
Gramática, Martín Nicolás; Pianzola, Guido; Alfici, Hugo David; García, Mario Alejandro
Año de publicación
2026
Idioma
inglés
Tipo de recurso
parte de libro
Estado
versión aceptada
Descripción
Automated analysis of Pap smear images can assist pathologists in cervical cancer screening by reducing workload while maintaining an interpretable diagnostic workflow. This work investigates the relationship between object detection and image classification in a two-stage deep learning approach for the classification of conventional Pap smear images as normal or altered. The first stage uses YOLOv7 to detect and classify individual cells, while the second stage aggregates the detection results through a simple decision rule based on the number of detected altered cells and their confidence scores. Using the CRIC dataset, three object detection models were evaluated, and their outputs were compared with the corresponding ground-truth detections. The effect of detection quality on image-level classification was then assessed using recall, false-positive rate, precision, accuracy, F-score, and ROC-AUC. Different numbers of altered cells and probability thresholds were also evaluated, while classifications were compared against the assessments of two expert pathologists to account for inter-observer variability. The results show that degradation in object detection quality leads to a degradation in classification performance, but the effect is substantially attenuated at the second stage. Moreover, adjusting the classification rule can improve overall performance and, in some cases, compensate for detection errors. The optimal configuration depends on the diagnostic criteria of the pathologist, highlighting the influence of inter-observer variability. These findings indicate that two-stage approaches can provide robustness to imperfect cell detection while retaining the interpretability associated with explicit cell-level predictions, potentially allowing reductions in annotation requirements without substantial losses in image-level classification quality.
Fil: García, Mario Alejandro. Universidad Tecnológica Nacional. Facultad Regional Córdoba. Grupo de Investigación en Inteligencia Artificial; Argentina.
Fil: Gramática, Martín Nicolás . Universidad Tecnológica Nacional. Facultad Regional Córdoba. Grupo de Investigación en Inteligencia Artificial; Argentina.
Fil: Pianzola, Guido. Hospital Inter zonal San Roque. Grupo de Investigación en Inteligencia Artificial; Argentina.
Fil: Pianzola, Guido. Hospital Inter zonal San Roque. Grupo de Investigación en Inteligencia Artificial; Argentina.
Fil: Alfici, Hugo David. Hospital Madariaga. Grupo de Investigación en Inteligencia Artificial; Argentina.
Peer Reviewed
Fuente
Libro Advances in Image Processing, Reliability, and Artificial Intelligence. Elsevier.
Materia
Inteligencia artificial
Aprendizaje automático
Patología digital
Pap test
Nivel de accesibilidad
acceso abierto
Condiciones de uso
Attribution-NonCommercial-NoDerivs 2.5 Argentina
Repositorio
Repositorio Institucional Abierto (UTN)
Institución
Universidad Tecnológica Nacional
OAI Identificador
oai:ria.utn.edu.ar:20.500.12272/15565
Acceder
id RIAUTN_6a0a6619bc9891652f475881a814621b
oai_identifier_str oai:ria.utn.edu.ar:20.500.12272/15565
network_acronym_str RIAUTN
repository_id_str a
network_name_str Repositorio Institucional Abierto (UTN)
spelling Relationship between object detection and image classification stages in pap testGramática, Martín NicolásPianzola, GuidoAlfici, Hugo DavidGarcía, Mario AlejandroInteligencia artificialAprendizaje automáticoPatología digitalPap testAutomated analysis of Pap smear images can assist pathologists in cervical cancer screening by reducing workload while maintaining an interpretable diagnostic workflow. This work investigates the relationship between object detection and image classification in a two-stage deep learning approach for the classification of conventional Pap smear images as normal or altered. The first stage uses YOLOv7 to detect and classify individual cells, while the second stage aggregates the detection results through a simple decision rule based on the number of detected altered cells and their confidence scores. Using the CRIC dataset, three object detection models were evaluated, and their outputs were compared with the corresponding ground-truth detections. The effect of detection quality on image-level classification was then assessed using recall, false-positive rate, precision, accuracy, F-score, and ROC-AUC. Different numbers of altered cells and probability thresholds were also evaluated, while classifications were compared against the assessments of two expert pathologists to account for inter-observer variability. The results show that degradation in object detection quality leads to a degradation in classification performance, but the effect is substantially attenuated at the second stage. Moreover, adjusting the classification rule can improve overall performance and, in some cases, compensate for detection errors. The optimal configuration depends on the diagnostic criteria of the pathologist, highlighting the influence of inter-observer variability. These findings indicate that two-stage approaches can provide robustness to imperfect cell detection while retaining the interpretability associated with explicit cell-level predictions, potentially allowing reductions in annotation requirements without substantial losses in image-level classification quality.Fil: García, Mario Alejandro. Universidad Tecnológica Nacional. Facultad Regional Córdoba. Grupo de Investigación en Inteligencia Artificial; Argentina.Fil: Gramática, Martín Nicolás . Universidad Tecnológica Nacional. Facultad Regional Córdoba. Grupo de Investigación en Inteligencia Artificial; Argentina.Fil: Pianzola, Guido. Hospital Inter zonal San Roque. Grupo de Investigación en Inteligencia Artificial; Argentina.Fil: Pianzola, Guido. Hospital Inter zonal San Roque. Grupo de Investigación en Inteligencia Artificial; Argentina.Fil: Alfici, Hugo David. Hospital Madariaga. Grupo de Investigación en Inteligencia Artificial; Argentina.Peer ReviewedUniversidad Tecnológica Nacional Regional Córdoba.2026-09-14T19:17:15Z2026-05-01info:eu-repo/semantics/bookPartinfo:eu-repo/semantics/acceptedVersionhttp://purl.org/coar/resource_type/c_3248info:ar-repo/semantics/parteDeLibropdfapplication/pdfGramática, M. N., Pianzola, G., Alfici, H. D., & García, M. A. (2026). Relationship between object detection and image classification stages in Pap test. In Advances in Image Processing, Reliability, and Artificial Intelligence (pp. 275-291). Elsevier.978-0-443-34266-0https://hdl.handle.net/20.500.12272/15565Libro Advances in Image Processing, Reliability, and Artificial Intelligence. Elsevier.reponame:Repositorio Institucional Abierto (UTN)instname:Universidad Tecnológica Nacionalenginfo:eu-repo/semantics/openAccessAttribution-NonCommercial-NoDerivs 2.5 Argentinahttp://creativecommons.org/licenses/by-nc-nd/2.5/ar/Gramática, Martín Nicolás; Pianzola, Guido; Alfici, Hugo David; García, Mario Alejandrohttps://creativecommons.org/licenses/by-nc-nd/4.0/2026-10-01T11:56:41Zoai:ria.utn.edu.ar:20.500.12272/15565instacron:UTNInstitucionalhttp://ria.utn.edu.ar/Universidad públicaNo correspondehttp://ria.utn.edu.ar/oaigestionria@rec.utn.edu.ar; fsuarez@rec.utn.edu.arArgentinaNo correspondeNo correspondeNo correspondeopendoar:a2026-10-01 11:56:41.842Repositorio Institucional Abierto (UTN) - Universidad Tecnológica Nacionalfalse
dc.title.none.fl_str_mv Relationship between object detection and image classification stages in pap test
title Relationship between object detection and image classification stages in pap test
spellingShingle Relationship between object detection and image classification stages in pap test
Gramática, Martín Nicolás
Inteligencia artificial
Aprendizaje automático
Patología digital
Pap test
title_short Relationship between object detection and image classification stages in pap test
title_full Relationship between object detection and image classification stages in pap test
title_fullStr Relationship between object detection and image classification stages in pap test
title_full_unstemmed Relationship between object detection and image classification stages in pap test
title_sort Relationship between object detection and image classification stages in pap test
dc.creator.none.fl_str_mv Gramática, Martín Nicolás
Pianzola, Guido
Alfici, Hugo David
García, Mario Alejandro
author Gramática, Martín Nicolás
author_facet Gramática, Martín Nicolás
Pianzola, Guido
Alfici, Hugo David
García, Mario Alejandro
author_role author
author2 Pianzola, Guido
Alfici, Hugo David
García, Mario Alejandro
author2_role author
author
author
dc.subject.none.fl_str_mv Inteligencia artificial
Aprendizaje automático
Patología digital
Pap test
topic Inteligencia artificial
Aprendizaje automático
Patología digital
Pap test
dc.description.none.fl_txt_mv Automated analysis of Pap smear images can assist pathologists in cervical cancer screening by reducing workload while maintaining an interpretable diagnostic workflow. This work investigates the relationship between object detection and image classification in a two-stage deep learning approach for the classification of conventional Pap smear images as normal or altered. The first stage uses YOLOv7 to detect and classify individual cells, while the second stage aggregates the detection results through a simple decision rule based on the number of detected altered cells and their confidence scores. Using the CRIC dataset, three object detection models were evaluated, and their outputs were compared with the corresponding ground-truth detections. The effect of detection quality on image-level classification was then assessed using recall, false-positive rate, precision, accuracy, F-score, and ROC-AUC. Different numbers of altered cells and probability thresholds were also evaluated, while classifications were compared against the assessments of two expert pathologists to account for inter-observer variability. The results show that degradation in object detection quality leads to a degradation in classification performance, but the effect is substantially attenuated at the second stage. Moreover, adjusting the classification rule can improve overall performance and, in some cases, compensate for detection errors. The optimal configuration depends on the diagnostic criteria of the pathologist, highlighting the influence of inter-observer variability. These findings indicate that two-stage approaches can provide robustness to imperfect cell detection while retaining the interpretability associated with explicit cell-level predictions, potentially allowing reductions in annotation requirements without substantial losses in image-level classification quality.
Fil: García, Mario Alejandro. Universidad Tecnológica Nacional. Facultad Regional Córdoba. Grupo de Investigación en Inteligencia Artificial; Argentina.
Fil: Gramática, Martín Nicolás . Universidad Tecnológica Nacional. Facultad Regional Córdoba. Grupo de Investigación en Inteligencia Artificial; Argentina.
Fil: Pianzola, Guido. Hospital Inter zonal San Roque. Grupo de Investigación en Inteligencia Artificial; Argentina.
Fil: Pianzola, Guido. Hospital Inter zonal San Roque. Grupo de Investigación en Inteligencia Artificial; Argentina.
Fil: Alfici, Hugo David. Hospital Madariaga. Grupo de Investigación en Inteligencia Artificial; Argentina.
Peer Reviewed
description Automated analysis of Pap smear images can assist pathologists in cervical cancer screening by reducing workload while maintaining an interpretable diagnostic workflow. This work investigates the relationship between object detection and image classification in a two-stage deep learning approach for the classification of conventional Pap smear images as normal or altered. The first stage uses YOLOv7 to detect and classify individual cells, while the second stage aggregates the detection results through a simple decision rule based on the number of detected altered cells and their confidence scores. Using the CRIC dataset, three object detection models were evaluated, and their outputs were compared with the corresponding ground-truth detections. The effect of detection quality on image-level classification was then assessed using recall, false-positive rate, precision, accuracy, F-score, and ROC-AUC. Different numbers of altered cells and probability thresholds were also evaluated, while classifications were compared against the assessments of two expert pathologists to account for inter-observer variability. The results show that degradation in object detection quality leads to a degradation in classification performance, but the effect is substantially attenuated at the second stage. Moreover, adjusting the classification rule can improve overall performance and, in some cases, compensate for detection errors. The optimal configuration depends on the diagnostic criteria of the pathologist, highlighting the influence of inter-observer variability. These findings indicate that two-stage approaches can provide robustness to imperfect cell detection while retaining the interpretability associated with explicit cell-level predictions, potentially allowing reductions in annotation requirements without substantial losses in image-level classification quality.
publishDate 2026
dc.date.none.fl_str_mv 2026-09-14T19:17:15Z
2026-05-01
dc.type.none.fl_str_mv info:eu-repo/semantics/bookPart
info:eu-repo/semantics/acceptedVersion
http://purl.org/coar/resource_type/c_3248
info:ar-repo/semantics/parteDeLibro
format bookPart
status_str acceptedVersion
dc.identifier.none.fl_str_mv Gramática, M. N., Pianzola, G., Alfici, H. D., & García, M. A. (2026). Relationship between object detection and image classification stages in Pap test. In Advances in Image Processing, Reliability, and Artificial Intelligence (pp. 275-291). Elsevier.
978-0-443-34266-0
https://hdl.handle.net/20.500.12272/15565
identifier_str_mv Gramática, M. N., Pianzola, G., Alfici, H. D., & García, M. A. (2026). Relationship between object detection and image classification stages in Pap test. In Advances in Image Processing, Reliability, and Artificial Intelligence (pp. 275-291). Elsevier.
978-0-443-34266-0
url https://hdl.handle.net/20.500.12272/15565
dc.language.none.fl_str_mv eng
language eng
dc.rights.none.fl_str_mv info:eu-repo/semantics/openAccess
Attribution-NonCommercial-NoDerivs 2.5 Argentina
http://creativecommons.org/licenses/by-nc-nd/2.5/ar/
Gramática, Martín Nicolás; Pianzola, Guido; Alfici, Hugo David; García, Mario Alejandro
https://creativecommons.org/licenses/by-nc-nd/4.0/
eu_rights_str_mv openAccess
rights_invalid_str_mv Attribution-NonCommercial-NoDerivs 2.5 Argentina
http://creativecommons.org/licenses/by-nc-nd/2.5/ar/
Gramática, Martín Nicolás; Pianzola, Guido; Alfici, Hugo David; García, Mario Alejandro
https://creativecommons.org/licenses/by-nc-nd/4.0/
dc.format.none.fl_str_mv pdf
application/pdf
dc.publisher.none.fl_str_mv Universidad Tecnológica Nacional Regional Córdoba.
publisher.none.fl_str_mv Universidad Tecnológica Nacional Regional Córdoba.
dc.source.none.fl_str_mv Libro Advances in Image Processing, Reliability, and Artificial Intelligence. Elsevier.
reponame:Repositorio Institucional Abierto (UTN)
instname:Universidad Tecnológica Nacional
reponame_str Repositorio Institucional Abierto (UTN)
collection Repositorio Institucional Abierto (UTN)
instname_str Universidad Tecnológica Nacional
repository.name.fl_str_mv Repositorio Institucional Abierto (UTN) - Universidad Tecnológica Nacional
repository.mail.fl_str_mv gestionria@rec.utn.edu.ar; fsuarez@rec.utn.edu.ar
_version_ 1877862346054959104
score 13.365483

Publicaciones similares