Your browser doesn't support javascript.
loading
Prediction of molecular subclasses of uveal melanoma by deep learning using routine haematoxylin-eosin-stained tissue slides.
Akram, Farhan; de Bruyn, Daniël P; van den Bosch, Quincy C C; Trandafir, Teodora E; van den Bosch, Thierry P P; Verdijk, Rob M; de Klein, Annelies; Kiliç, Emine; Stubbs, Andrew P; Brosens, Erwin; von der Thüsen, Jan H.
Afiliação
  • Akram F; Department of Pathology and Clinical Bioinformatics, Erasmus MC Rotterdam, Rotterdam, the Netherlands.
  • de Bruyn DP; Ophthalmology, Erasmus MC Rotterdam, Rotterdam, the Netherlands.
  • van den Bosch QCC; Clinical Genetics, Erasmus MC Rotterdam, Rotterdam, the Netherlands.
  • Trandafir TE; Cancer Institute, Erasmus MC Rotterdam, Rotterdam, the Netherlands.
  • van den Bosch TPP; Ophthalmology, Erasmus MC Rotterdam, Rotterdam, the Netherlands.
  • Verdijk RM; Clinical Genetics, Erasmus MC Rotterdam, Rotterdam, the Netherlands.
  • de Klein A; Cancer Institute, Erasmus MC Rotterdam, Rotterdam, the Netherlands.
  • Kiliç E; Department of Pathology and Clinical Bioinformatics, Erasmus MC Rotterdam, Rotterdam, the Netherlands.
  • Stubbs AP; Department of Pathology and Clinical Bioinformatics, Erasmus MC Rotterdam, Rotterdam, the Netherlands.
  • Brosens E; Department of Pathology and Clinical Bioinformatics, Erasmus MC Rotterdam, Rotterdam, the Netherlands.
  • von der Thüsen JH; Ophthalmology, Erasmus MC Rotterdam, Rotterdam, the Netherlands.
Histopathology ; 2024 Jul 01.
Article em En | MEDLINE | ID: mdl-38952117
ABSTRACT

AIMS:

Uveal melanoma has a high propensity to metastasize. Prognosis is associated with specific driver mutations and copy number variations, and these can only be obtained after genetic testing. In this study we evaluated the efficacy of patient outcome prediction using deep learning on haematoxylin and eosin (HE)-stained primary uveal melanoma slides in comparison to molecular testing.

METHODS:

In this retrospective study of patients with uveal melanoma, 113 patients from the Erasmus Medical Centre who underwent enucleation had tumour tissue analysed for molecular classification between 1993 and 2020. Routine HE-stained slides were scanned to obtain whole-slide images (WSI). After annotation of regions of interest, tiles of 1024 × 1024 pixels were extracted at a magnification of 40×. An ablation study to select the best-performing deep-learning model was carried out using three state-of-the-art deep-learning models (EfficientNet, Vision Transformer, and Swin Transformer).

RESULTS:

Deep-learning models were subjected to a training cohort (n = 40), followed by a validation cohort (n = 20), and finally underwent a test cohort (n = 48). A k-fold cross-validation (k = 3) of validation and test cohorts (n = 113 of three classes BAP1, SF3B1, EIF1AX) demonstrated Swin Transformer as the best-performing deep-learning model to predict molecular subclasses based on HE stains. The model achieved an accuracy of 0.83 ± 0.09 on the validation cohort and 0.75 ± 0.04 on the test cohort. Within the subclasses, this model correctly predicted 70% BAP1-mutated, 61% SF3B1-mutated and 80% EIF1AX-mutated UM in the test set.

CONCLUSIONS:

This study showcases the potential of the deep-learning methodology for predicting molecular subclasses in a multiclass manner using HE-stained WSI. This development holds promise for advanced prognostication of UM patients without the need of molecular or immunohistochemical testing. Additionally, this study suggests there are distinct histopathological features per subclass; mainly utilizing epithelioid cellular morphology for BAP1-classification, but an unknown feature distinguishes EIF1AX and SF3B1.

Texto completo: 1 Coleções: 01-internacional Base de dados: MEDLINE Idioma: En Revista: Histopathology Ano de publicação: 2024 Tipo de documento: Article País de afiliação: Holanda

Texto completo: 1 Coleções: 01-internacional Base de dados: MEDLINE Idioma: En Revista: Histopathology Ano de publicação: 2024 Tipo de documento: Article País de afiliação: Holanda