Fine-tuned deep neural networks for polyp detection in colonoscopy images

• Published in: Personal and ubiquitous computing 2023-04, Vol.27 (2), p.235-247
• Main Authors: Ellahyani, Ayoub, Jaafari, Ilyas El, Charfi, Said, Ansari, Mohamed El
• Format: Article
• Language: English
• Subjects: Abnormalities; Artificial neural networks; Cancer; Colonoscopy; Colorectal cancer; Computer Science; Datasets; Diagnostic systems; Medical imaging; Mobile Computing; Neural networks; Original Paper; Personal Computing; Polyps; User Interfaces and Human Computer Interaction
• ISSN: 1617-4909; 1617-4917
• DOI: 10.1007/s00779-021-01660-y

Colorectal cancer is one of the leading causes of cancer death worldwide. It can appear in different forms depending on its location, and in most cases, these are tumors called polyps. A diagnostic model which is built can aid doctors but can not replace them. Thus, automated diagnostic technics to detect symptoms of illness or abnormalities in video colonoscopy or wireless capsule endoscopy (WCE) are adopted as an excellent enhancement technique for doctors. In this work, a new computer-assisted diagnosis method for polyp detection is proposed. After a preprocessing step, a fusion of two deep neural networks (DNNs), which were pre-trained on millions of labeled natural images (ImageNet), are fine-tuned and used to extract deep features to perform the polyp detection. The fine-tuning is employed using the Kvasir-seg dataset images. Moreover, the weights of the initial layers of the networks used in this work are frozen. Finally, we concatenated the fully connected outputs of the fine-tuned models to perform the binary classification. The proposed method achieved 0.919, 0.897, and 0.907 on the CVC-ClinicDB dataset, and 0.876, 0.910, and 0.893 on the ETIS-LaribPolypDB dataset in terms of precision, recall, and F-measure metrics, respectively. The results obtained are satisfactory when compared to the state-of-the-art methods.