A Custom-Based Deep Learning Architecture for Classifying Colorectal Polyps into Premalignant and Benign

Diogen Babuc

Abstract.

This paper explores specific deep learning architectures for binary classification of colorectal polyps and takes into consideration their risk assessment for malignancy. Relying on deep neural models, virtual biopsy aims to replace histopathologic diagnosis during colonoscopy for the detection and visual classification of colorectal polyps. The ZF NET architecture is the most accurate (over 84%), while ResNet stands out in effectively indicating patients with premalignant colorectal polyps (over 92% recall). The advantage of the proposed model, Bionnica, stems from personalized improvement with a rule-based layer, guiding the learning process and supporting medical staff in decision-making regarding the polyps' condition. According to evaluation metrics, Bionnica outperforms all architectures engaged in this experiment, for the selected dataset.