Globally, colorectal cancer is one of the leading causes of mortality. Colonoscopies and the early removal of polys significantly increase the survival rate of this cancer, but this intervention depends on the accurate detection of polys in the surrounding tissues. Missing a poly has serious consequences. One way to guard against human error is to develop automatic polyp detection systems. Deep learning semantic segmentation offers one approach to solving the problem of poly detection. In this work, we propose an ensemble of ensembles composed of two deep convolutional neural networks (DCNNs): DeepLabV3+ and HarDNet. Diversity among the single classifiers is enforced on the data level using different data augmentation approaches and on the classifier level with two DCNNs: DeepLabV3+ and HardNet, each using an encoder-decoder unit. In addition, ensembles of DeepLabV3+ are built using fifteen loss functions. Our best ensembles are tested on a large dataset composed of samples taken from five polyp benchmarks. Ensembles are assessed and compared with the best method reported in the literature and shown to produce state-of-the-art results. The source code, the dataset, and the testing protocol used in this study are freely available at https://github.com/LorisNanni.
Polyp Segmentation with Deep Ensembles and Data Augmentation
Loreggia A.;
2023-01-01
Abstract
Globally, colorectal cancer is one of the leading causes of mortality. Colonoscopies and the early removal of polys significantly increase the survival rate of this cancer, but this intervention depends on the accurate detection of polys in the surrounding tissues. Missing a poly has serious consequences. One way to guard against human error is to develop automatic polyp detection systems. Deep learning semantic segmentation offers one approach to solving the problem of poly detection. In this work, we propose an ensemble of ensembles composed of two deep convolutional neural networks (DCNNs): DeepLabV3+ and HarDNet. Diversity among the single classifiers is enforced on the data level using different data augmentation approaches and on the classifier level with two DCNNs: DeepLabV3+ and HardNet, each using an encoder-decoder unit. In addition, ensembles of DeepLabV3+ are built using fifteen loss functions. Our best ensembles are tested on a large dataset composed of samples taken from five polyp benchmarks. Ensembles are assessed and compared with the best method reported in the literature and shown to produce state-of-the-art results. The source code, the dataset, and the testing protocol used in this study are freely available at https://github.com/LorisNanni.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
