Enhancing Generalizability of Deep Learning Polyp Segmentation Using Online Spatial Interpolation and Hue Transformation

Mahmood Haithami*, Amr Ahmed, Iman Yi Liao

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference proceeding (ISBN)peer-review

Abstract

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.

Original languageEnglish
Title of host publicationAdvances in Brain Inspired Cognitive Systems - 13th International Conference, BICS 2023, Proceedings
EditorsJinchang Ren, Amir Hussain, Iman Yi Liao, Rongjun Chen, Kaizhu Huang, Huimin Zhao, Xiaoyong Liu, Ping Ma, Thomas Maul
PublisherSpringer Science and Business Media Deutschland GmbH
Pages41-50
Number of pages10
ISBN (Print)9789819714162
DOIs
Publication statusPublished - 22 May 2024
Event13th International Conference on Brain Inspired Cognitive Systems, BICS 2023 - Kuala Lumpur, Malaysia
Duration: 5 Aug 20236 Aug 2023

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume14374 LNAI

Conference

Conference13th International Conference on Brain Inspired Cognitive Systems, BICS 2023
Country/TerritoryMalaysia
CityKuala Lumpur
Period5/08/236/08/23

Keywords

  • Deep learning
  • Generalizability
  • Polyp Segmentation
  • Spatial interpolation

Fingerprint

Dive into the research topics of 'Enhancing Generalizability of Deep Learning Polyp Segmentation Using Online Spatial Interpolation and Hue Transformation'. Together they form a unique fingerprint.

Cite this