TY - GEN
T1 - Enhancing Generalizability of Deep Learning Polyp Segmentation Using Online Spatial Interpolation and Hue Transformation
AU - Haithami, Mahmood
AU - Ahmed, Amr
AU - Liao, Iman Yi
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024.
PY - 2024/5/22
Y1 - 2024/5/22
N2 - Polyps, which are precursors to colon cancer, can be detected early to reduce mortality rates. However, the limited availability of public datasets and the variability of polyp shapes, textures, and colors restrict the generalizability of existing deep learning models. To overcome this challenge, researchers often employ data augmentation techniques or generative models to increase the number of training samples, regardless of the downstream learning task (i.e., polyp segmentation). In this study, we propose a deep learning framework that combines an image transformation layer with a segmentation model, where the transformed images serve as input for the segmentation model. The image transformation layer comprises a random hue shifting function and an autoencoder. The autoencoder removes textures while preserving key polyp features, transforming the input images. To control the intensity of the transformation, we employ a simple interpolation between the original and transformed images. During training, the image transformation layer generates multiple levels of texture and color variations for each input image in every epoch, effectively regularizing the segmentation model. By exposing the segmentation model to different texture and color levels within the same training batch, we encourage the model to update its weights based on intrinsic features present in both the original images and their corresponding transformed versions. This approach enhances the generalizability of deep learning models on unseen test sets. Experimental results using various configurations consistently demonstrate significant improvements in polyp Intersection over Union (IoU) ranging from 1.8% to 16.4% across different test sets.
AB - Polyps, which are precursors to colon cancer, can be detected early to reduce mortality rates. However, the limited availability of public datasets and the variability of polyp shapes, textures, and colors restrict the generalizability of existing deep learning models. To overcome this challenge, researchers often employ data augmentation techniques or generative models to increase the number of training samples, regardless of the downstream learning task (i.e., polyp segmentation). In this study, we propose a deep learning framework that combines an image transformation layer with a segmentation model, where the transformed images serve as input for the segmentation model. The image transformation layer comprises a random hue shifting function and an autoencoder. The autoencoder removes textures while preserving key polyp features, transforming the input images. To control the intensity of the transformation, we employ a simple interpolation between the original and transformed images. During training, the image transformation layer generates multiple levels of texture and color variations for each input image in every epoch, effectively regularizing the segmentation model. By exposing the segmentation model to different texture and color levels within the same training batch, we encourage the model to update its weights based on intrinsic features present in both the original images and their corresponding transformed versions. This approach enhances the generalizability of deep learning models on unseen test sets. Experimental results using various configurations consistently demonstrate significant improvements in polyp Intersection over Union (IoU) ranging from 1.8% to 16.4% across different test sets.
KW - Deep learning
KW - Generalizability
KW - Polyp Segmentation
KW - Spatial interpolation
UR - http://www.scopus.com/inward/record.url?scp=85195106928&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85195106928&partnerID=8YFLogxK
UR - https://www.mendeley.com/catalogue/a9fb7d3e-c9b9-34e9-9ddf-9c0db67d2726/
U2 - 10.1007/978-981-97-1417-9_4
DO - 10.1007/978-981-97-1417-9_4
M3 - Conference proceeding (ISBN)
AN - SCOPUS:85195106928
SN - 9789819714162
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 41
EP - 50
BT - Advances in Brain Inspired Cognitive Systems - 13th International Conference, BICS 2023, Proceedings
A2 - Ren, Jinchang
A2 - Hussain, Amir
A2 - Liao, Iman Yi
A2 - Chen, Rongjun
A2 - Huang, Kaizhu
A2 - Zhao, Huimin
A2 - Liu, Xiaoyong
A2 - Ma, Ping
A2 - Maul, Thomas
PB - Springer Science and Business Media Deutschland GmbH
T2 - 13th International Conference on Brain Inspired Cognitive Systems, BICS 2023
Y2 - 5 August 2023 through 6 August 2023
ER -