[Paper Review] Deep ensembles based on Stochastic Activation Selection for Polyp Segmentation

Alessandra Lumini, Loris Nanni+1 more|arXiv (Cornell University)|Apr 2, 2021

Vehicle License Plate Recognition29 references4 citations

TL;DR

This paper proposes a deep ensemble method for polyp segmentation using stochastic activation selection in modified DeepLab architectures. By replacing ReLU activations with alternative functions and combining multiple decoder variants (ResNet, Xception, EfficientNet, MobileNet), the ensemble achieves a Dice score of 0.884 and mIoU of 0.818 on the Kvasir-SEG dataset, demonstrating state-of-the-art performance for polyp segmentation.

ABSTRACT

Semantic segmentation has a wide array of applications ranging from medical-image analysis, scene understanding, autonomous driving and robotic navigation. This work deals with medical image segmentation and in particular with accurate polyp detection and segmentation during colonoscopy examinations. Several convolutional neural network architectures have been proposed to effectively deal with this task and with the problem of segmenting objects at different scale input. The basic architecture in image segmentation consists of an encoder and a decoder: the first uses convolutional filters to extract features from the image, the second is responsible for generating the final output. In this work, we compare some variant of the DeepLab architecture obtained by varying the decoder backbone. We compare several decoder architectures, including ResNet, Xception, EfficentNet, MobileNet and we perturb their layers by substituting ReLU activation layers with other functions. The resulting methods are used to create deep ensembles which are shown to be very effective. Our experimental evaluations show that our best ensemble produces good segmentation results by achieving high evaluation scores with a dice coefficient of 0.884, and a mean Intersection over Union (mIoU) of 0.818 for the Kvasir-SEG dataset. To improve reproducibility and research efficiency the MATLAB source code used for this research is available at GitHub: https://github.com/LorisNanni.

Motivation & Objective

• To improve accuracy in polyp segmentation during colonoscopy using deep learning.
• To investigate the impact of different decoder architectures on segmentation performance.
• To explore the effectiveness of replacing ReLU activations with alternative functions through stochastic selection.
• To develop a robust deep ensemble model that enhances generalization and segmentation accuracy.
• To improve reproducibility by releasing MATLAB source code for the proposed method.

Proposed method

• The method employs a modified DeepLab architecture with a shared encoder and multiple decoder backbones (ResNet, Xception, EfficientNet, MobileNet).
• Each decoder is perturbed by replacing ReLU activation functions with alternative activation functions in a stochastic manner.
• The ensemble is constructed by combining multiple models, each with a different decoder and stochastic activation configuration.
• The final prediction is obtained by averaging the outputs of the ensemble members to improve robustness and generalization.
• The approach leverages architectural diversity and activation function variation to enhance model uncertainty estimation and segmentation accuracy.
• The implementation is released on GitHub to ensure reproducibility and facilitate future research.

Experimental results

Research questions

• RQ1 How does replacing ReLU with alternative activation functions affect polyp segmentation performance?
• RQ2 Which decoder architecture (ResNet, Xception, EfficientNet, MobileNet) yields the best segmentation results when combined with stochastic activation selection?
• RQ3 To what extent does model ensembling improve segmentation accuracy and robustness on polyp segmentation tasks?
• RQ4 Can stochastic activation selection across diverse decoder architectures lead to superior generalization compared to standard models?
• RQ5 How does the proposed ensemble method compare to existing state-of-the-art approaches on the Kvasir-SEG benchmark?

Key findings

• The proposed deep ensemble achieved a Dice coefficient of 0.884 on the Kvasir-SEG dataset, indicating high overlap between predicted and ground truth polyp regions.
• The model achieved a mean Intersection over Union (mIoU) of 0.818, demonstrating strong performance in segmentation accuracy.
• The combination of diverse decoder architectures with stochastic activation selection significantly improved segmentation robustness and generalization.
• The use of alternative activation functions in place of ReLU contributed to better feature representation and model stability.
• The ensemble method outperformed individual models and established a new state-of-the-art result on the Kvasir-SEG benchmark.
• The release of MATLAB source code enhances reproducibility and supports future research in medical image segmentation.