Rethinking the Transfer Learning for FCN Based Polyp Segmentation in Colonoscopy

Abstract

Besides the complex nature of colonoscopy frames with intrinsic frame formation artefacts such as light reflections and the diversity of polyp types/shapes, the publicly available polyp segmentation training datasets are limited, small and imbalanced. In this case, the automated polyp segmentation using a deep neural network remains an open challenge due to the overfitting of training on small datasets. We proposed a simple yet effective polyp segmentation pipeline that couples the segmentation (FCN) and classification (CNN) tasks. We find the effectiveness of interactive weight transfer between dense and coarse vision tasks that mitigates the overfitting in learning. This motivates us to design a new training scheme within our segmentation pipeline. Our method is evaluated on CVC-EndoSceneStill and Kvasir-SEG datasets. It achieves 4.34% and 5.70% Polyp-IoU improvements compared to the state-of-the-art methods on the EndoSceneStill and Kvasir-SEG datasets, respectively and achieves real-time performance in inference. The model and code are available at: https://github.com/MELSunny/Keras-FCN INDEX TERMS Colonoscopy, real-time polyp segmentation, transfer learning, convolutional neural network

Automated Colorectal Polyps Detection from Endoscopic Images using MultiResUNet Framework with Attention Guided Segmentation

“…As a result, despite having fewer parameters, the proposed system is still able to maintain high performance. The DSC score of FCN + CNN was 0.8922, but the IoU score was 0.8022; the proposed model received 0.8277 IoU, an improvement of 3.02% [19]. The DDANet had the highest recall score, but the IoU, DSC, and precision metrics needed improvement [21].…”

“…CAD systems have been used by many researchers to develop an automated procedure for identifying polyps. Many researchers trained their models using the Kvasir-SEG dataset, while others worked on post-processing techniques to make the models robust [2,[16][17][18][19][20][21][22][23][24][25][26]. The segmentation of endoscopic images based on semantic information has been extensively studied in medical imaging [27].…”

IRv2-Net: A Deep Learning Framework for Enhanced Polyp Segmentation Performance Integrating InceptionResNetV2 and UNet Architecture with Test Time Augmentation Techniques

“…However, contemporary methods increasingly make use of CNN and pre-trained networks. The Kvasir-SEG dataset [ 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 ], the CVC-ClinicDB dataset [ 24 , 25 , 26 , 27 , 28 , 29 , 30 ], and the TTA [ 31 , 32 , 33 , 34 , 35 , 36 , 37 ] are frequently used datasets in analyses of polyp detection and segmentation networks.…”

“…Impressively, it achieved a processing speed of 54.60, outperforming its predecessor. Wen et al [ 19 ] devised a highly effective solution to tackle overfitting by combining FCN and CNN. Their revolutionary segmentation pipeline training method included an interactive weight transfer technique, which proved far superior to other approaches, delivering a remarkable DSC score of 0.8922.…”

Colorectal Polyp Detection and Comparative Evaluation Based on Deep Learning Approaches

“…However, the colonoscopy examination is time-consuming and highly labor-intensive. It is common for polyps to be overlooked during a colonoscopy procedure due to their small size and visual characteristics [6]. Moreover, the endoscopist may make an error in diagnosis due to eyestrain or lack of concentration.…”