A DWT-based encoder-decoder network for Specularity segmentation in colonoscopy images

Specularity segmentation in colonoscopy images is a crucial pre-processing step for efficient computational diagnosis. The presence of these specular highlights could mislead the detectors that are intended towards precise identification of biomarkers. Conventional methods adopted so far do not provide satisfactory results, especially in the overexposed regions. The potential of deep learning methods is still unexplored in the related problem domain. Our work aims at providing a solution for more accurate highlights segmentation to assist surgeons. In this paper, we propose a novel deep learning based approach that performs segmentation following a multi-resolution analysis. This is achieved by introducing discrete wavelet transform (DWT) into the proposed model. We replace the standard pooling layers with DWTs, which helps preserve information and circumvent the effect of overexposed regions. All analytical experiments are performed using a publicly available benchmark dataset, and an F1-score (%) of 83.10 ± 0.14 is obtained on the test set. The experimental results show that this technique outperforms state-of-the-art methods and performs significantly better in overexposed regions. The proposed model also performed superior to some deep learning models (but applied in different domains) when tested with our problem specifications. Our method provides segmentation outcomes that are closer to the actual segmentation done by experts. This ensures improved pre-processed colonoscopy images that aid in better diagnosis of colorectal cancer.