PROMETEO: A CNN-based computer-aided diagnosis system for WSI prostate cancer detection

Prostate cancer is currently one of the most commonly-diagnosed types of cancer among males. Although its death rate has dropped in the last decades, it is still a major concern and one of the leading causes of cancer death. Prostate biopsy is a test that confirms or excludes the presence of cancer...

Descripción completa

Detalles Bibliográficos
Autores: Durán López, Lourdes, Domínguez Morales, Juan Pedro, Conde Martín, A.F., Vicente Díaz, Saturnino, Linares Barranco, Alejandro
Tipo de recurso: artículo
Estado:Versión aceptada para publicación
Fecha de publicación:2020
País:España
Institución:Universidad de Sevilla (US)
Repositorio:idUS. Depósito de Investigación de la Universidad de Sevilla
OAI Identifier:oai:idus.us.es:11441/106950
Acceso en línea:https://hdl.handle.net/11441/106950
https://doi.org/10.1109/ACCESS.2020.3008868
Access Level:acceso abierto
Palabra clave:Convolutional Neural Networks (CNN)
Computer-aided diagnosis
Deep learning
Medical Image Analysis
Prostate cancer
Whole-slide images
Descripción
Sumario:Prostate cancer is currently one of the most commonly-diagnosed types of cancer among males. Although its death rate has dropped in the last decades, it is still a major concern and one of the leading causes of cancer death. Prostate biopsy is a test that confirms or excludes the presence of cancer in the tissue. Samples extracted from biopsies are processed and digitized, obtaining gigapixel-resolution images called wholeslide images, which are analyzed by pathologists. Automated intelligent systems could be useful for helping pathologists in this analysis, reducing fatigue and making the routine process faster. In this work, a novel Deep Learning based computer-aided diagnosis system is presented. This system is able to analyze wholeslide histology images that are first patch-sampled and preprocessed using different filters, including a novel patch-scoring algorithm that removes worthless areas from the tissue. Then, patches are used as input to a custom Convolutional Neural Network, which gives a report showing malignant regions on a heatmap. The impact of applying a stain-normalization process to the patches is also analyzed in order to reduce color variability between different scanners. After training the network with a 3-fold cross-validation method, 99.98% accuracy, 99.98% F1 score and 0.999 AUC are achieved on a separate test set. The computation time needed to obtain the heatmap of a whole-slide image is, on average, around 15 s. Our custom network outperforms other state-of-the-art works in terms of computational complexity for a binary classification task between normal and malignant prostate whole-slide images at patch level.