Objective: One of the fundamental and crucial tasks for the automated diagnosis of colorectal cancer is the segmentation of the acute gastrointestinal lesions, most commonly colorectal polyps. Therefore, in this work, we present a novel lightweight encoder-decoder mode of architecture with the attention mechanism to address this challenging task.
Methods: The proposed Li-SegPNet architecture harnesses cross-dimensional interaction in feature maps with novel encoder block with modified triplet attention. We have used atrous spatial pyramid pooling to handle the problem of segmenting objects at multiple scales. We also address the semantic gap between the encoder and decoder through a modified skip connection using attention gating.
Results: We applied our model to colonoscopy still images and trained and validated it on two publicly available datasets, Kvasir-SEG and CVC-ClinicDB. We achieve mean Intersection-Over-Union (mIoU) and dice scores of 0.88, 0.9058 and 0.8969, 0.9372 on Kvasir-SEG and CVC-ClinicDB, respectively. We analyze the generalizability of Li-SegPNet by testing it on two independent previously unseen datasets, Hyper-Kvasir and EndoTect 2020, and establish the model efficiency in cross-dataset evaluation. We employ multi-scale testing to examine the model performance on different sizes of polyps. Li-SegPNet performs best on medium-sized polyps with a mIoU and dice score of 0.9086 and 0.9137, respectively on the Kvasir-SEG dataset and 0.9425, 0.9434 of mIoU and dice score, respectively on CVC-ClinicDB.
Conclusion: The experimental results convey that we establish a new benchmark on these four datasets for the segmentation of polyps.
Significance: The proposed model can be used as a new benchmark model for polyps segmentation. Lesser parameters in comparison to other models give the edge in the applicability of the proposed Li-SegPNet model in real-time clinical analysis.