AttenUNeT X with iterative feedback mechanisms for robust deep learning skin lesion segmentation

TL;DR

This paper introduces AttenUNeT X, an improved U-Net model for skin lesion segmentation that achieves high accuracy using feedback mechanisms and attention modules.

Contribution

The novel integration of a feedback mechanism, Order Statistics Layer, and enhanced attention modules in a U-Net architecture for skin lesion segmentation.

Findings

AttenUNeT X achieved a Dice coefficient of 0.9211 on the ISIC 2018 dataset.

The model demonstrated consistent performance across multiple datasets including PH2 and ISIC 2017.

The model's enhancements improved boundary precision and contextual learning for skin lesion segmentation.

Abstract

Accurate skin lesion segmentation is critical for improving early diagnosis of skin cancer. In this study, we propose AttenUNeT X, a novel extension of the U-Net architecture that integrates three key enhancements: (i) a feedback mechanism within decoder blocks to iteratively refine spatial features, (ii) a custom Order Statistics Layer (OSL) to capture extreme-value lesion patterns, and (iii) enhanced attention modules to prioritize diagnostically relevant regions. These progresses improve segmentation performance by allowing for a targeted reaction to important lesion features. These additions work in synergy to improve boundary precision and contextual learning. The model was trained and validated using the International Skin Imaging Collaboration (ISIC 2018) Dataset, with PH2-Pedro Hispano Hospital Dataset was experimented for comparative analysis and ISIC 2017 used for external testing and cross-validation, respectively and guarantee dependable performance across a range of images distributions. Our preprocessing pipeline included hair removal, resizing, normalization, and extensive data augmentation to promote robustness. To facilitate model generalizability, the preprocessing pipeline featured data augmentation and images enhancement, and the attention-augmented encoder-decoder layers of the design highlighted key lesion features. Experimental results demonstrate strong performance, achieving a Dice coefficient of 0.9211, Intersection over Union (IoU) of 0.8533, and pixel accuracy of 0.9824 on ISIC 2018. Similarly high metrics were observed on PH2 and ISIC 2017 Datasets. These outcomes validate the proposed model’s potential for reliable deployment in clinical dermatological workflows.

Figures