Serviços Personalizados
Journal
Artigo
Indicadores
- Citado por SciELO
- Acessos
Links relacionados
- Citado por Google
- Similares em SciELO
- Similares em Google
Compartilhar
TecnoLógicas
versão impressa ISSN 0123-7799versão On-line ISSN 2256-5337
Resumo
GOMEZ, Andrés; LEON-PEREZ, Fabián; PLAZAS-WADYNSKI, Miguel e MARTINEZ-CARRILLO, Fabio. Multilevel Segmentation of Gleason Patterns using Convolutional Representations in Histopathological Images. TecnoL. [online]. 2021, vol.24, n.52, pp.176-196. Epub 14-Fev-2022. ISSN 0123-7799. https://doi.org/10.22430/22565337.2132.
The Gleason score is the most widely used grading system to diagnose and quantify the aggressiveness of prostate cancer, stratifying regional abnormal patterns on histological images. Nonetheless, recent studies into the Gleason score have reported moderate concordance values of 0.55 (kappa value) in the diagnosis of the disease. This study introduces a convolutional representation for the semantic segmentation and stratification of regions in histological images implementing the Gleason score and three levels of representation. On the first level, a regional network of the Mask R-CNN type is trained with complete annotations to define regional delineations, being effective in locations with general structures. On the second level, using the same architecture, a model is trained only with overlapping annotations from the first scheme, which are difficult-to-classify regions. Finally, a third level of representation produces a more granular description of the regions, considering the regions resulting from the activations of the first level. The final segmentation results from the superposition of the three levels of representation. The proposed strategy was validated and trained on a public set with 886 histological images. The segmentations thus generated achieved an average Area Under the Precision-Recall Curve (AUPRC) of 0.8 ± 0.18 and 0.76 ± 0.15 regarding the diagnoses of two pathologists, respectively. The results show regional intersection levels close to those of the reference pathologists. The proposed strategy is a potential tool to be implemented in clinical support and analysis.
Palavras-chave : Semantic segmentation; deep learning; Gleason score; histopathological images; prostate cancer.