Colorectal Polyp Segmentation: Impact of Combining Different Datasets on Deep Learning Model Performance

Araujo, Sandro Luis de; Scheeren, Michel Hanzen; Aguiar, Rubens Miguel Gomes; Mendes, Eduardo; Franco, Ricardo Augusto Pereira; Paula Filho, Pedro Luiz de

http://hdl.handle.net/10198/34428

Use this identifier to reference this record.

Name:	Description:	Size:	Format:
Colorectal Polyp.pdf		3.58 MB	Adobe PDF	Download

Send Feedback

Authors

Araujo, Sandro Luis de

Scheeren, Michel Hanzen

Aguiar, Rubens Miguel Gomes

Mendes, Eduardo

Franco, Ricardo Augusto Pereira

Paula Filho, Pedro Luiz de

Abstract(s)

Colorectal cancer is a major health concern, ranking as one of the most common and deadly forms of cancer. It typically begins as polyps, which are abnormal growths in the intestinal mucosa. Identifying and removing these polyps through colonoscopy is a crucial preventative measure. However, even experienced professionals can overlook some polyps during examinations. In this context, segmentation algorithms can assist medical professionals by identifying areas in an image that correspond to a polyp. These algorithms, which rely on deep learning, require extensive image datasets to effectively learn how to identify and segment polyps. This study aimed to identify public colonoscopy image datasets that contain polyps and to examine how combining these datasets might affect the performance of a deep learning-based segmentation algorithm. After selecting the datasets and defining their combinations, we trained a segmentation algorithm on each combination. The evaluation of the trained models showed that merging datasets can enhance model generalization, with increases of up to 0.242 in the dice coefficient and 0.256 in the Intersection over Union (IoU). These improvements could lead to higher diagnostic accuracy in clinical settings, enhancing efforts to prevent colorectal cancer.