[Paper Review] DONet: Dual Objective Networks for Skin Lesion Segmentation

Yaxiong Wang, Yunchao Wei+3 more|arXiv (Cornell University)|Aug 19, 2020

Cutaneous Melanoma Detection and ManagementMedicine50 references18 citations

TL;DR

This paper proposes DONet, a dual objective network for skin lesion segmentation that uses two symmetric decoders optimized with different loss functions to generate complementary probability maps, improving segmentation reliability. The method achieves state-of-the-art Dice scores of 0.881 on ISIC 2018 and 0.931 on PH² by integrating predictions from distinct optimization objectives and employing a recurrent context encoding module (RCEM) for multi-scale contextual feature learning.

ABSTRACT

Skin lesion segmentation is a crucial step in the computer-aided diagnosis of dermoscopic images. In the last few years, deep learning based semantic segmentation methods have significantly advanced the skin lesion segmentation results. However, the current performance is still unsatisfactory due to some challenging factors such as large variety of lesion scale and ambiguous difference between lesion region and background. In this paper, we propose a simple yet effective framework, named Dual Objective Networks (DONet), to improve the skin lesion segmentation. Our DONet adopts two symmetric decoders to produce different predictions for approaching different objectives. Concretely, the two objectives are actually defined by different loss functions. In this way, the two decoders are encouraged to produce differentiated probability maps to match different optimization targets, resulting in complementary predictions accordingly. The complementary information learned by these two objectives are further aggregated together to make the final prediction, by which the uncertainty existing in segmentation maps can be significantly alleviated. Besides, to address the challenge of large variety of lesion scales and shapes in dermoscopic images, we additionally propose a recurrent context encoding module (RCEM) to model the complex correlation among skin lesions, where the features with different scale contexts are efficiently integrated to form a more robust representation. Extensive experiments on two popular benchmarks well demonstrate the effectiveness of the proposed DONet. In particular, our DONet achieves 0.881 and 0.931 dice score on ISIC 2018 and $ ext{PH}^2$, respectively. Code will be made public available.

Motivation & Objective

• To address the challenge of high uncertainty and ambiguity in skin lesion segmentation due to variable lesion scale and weak lesion-background boundaries.
• To improve segmentation reliability by leveraging complementary predictions from different optimization objectives, rather than relying on a single loss function.
• To enhance multi-scale contextual representation for lesions of varying shapes and sizes through a novel recurrent context encoding module (RCEM).
• To achieve state-of-the-art performance on benchmark datasets by jointly optimizing for multiple segmentation objectives through a dual-decoder architecture.

Proposed method

• DONet employs a shared encoder followed by two symmetric decoders, each optimized with a different loss function (e.g., Dice loss and Focal Tversky loss), enabling distinct optimization objectives.
• The two decoders produce separate probability maps that are complementary, as each is biased toward optimizing its respective loss target.
• A joint decision mechanism aggregates the two probability maps to produce a final, more reliable segmentation map, reducing uncertainty.
• The recurrent context encoding module (RCEM) uses a convolutional LSTM to iteratively accumulate multi-scale contextual features from dilated convolutions with rates 1, 2, 4, and 8.
• The RCEM integrates features across different receptive fields through a recurrent mechanism, improving robustness to lesion scale variations.
• A multi-scale skip connection transfers contextual knowledge from the RCEM to the decoding paths, enhancing feature representation in the decoder.

Experimental results

Research questions

• RQ1 Can training two decoders with different loss functions improve segmentation reliability by generating complementary predictions?
• RQ2 How can multi-scale contextual features be effectively aggregated to improve robustness to lesion scale and shape variations?
• RQ3 Does a recurrent mechanism for context encoding outperform simple feature concatenation in capturing long-range dependencies in dermoscopic images?
• RQ4 To what extent do the dual objective architecture and RCEM contribute individually and jointly to segmentation performance?

Key findings

• DONet achieves a Dice score of 0.881 on the ISIC 2018 dataset, outperforming the baseline by 0.027 points.
• On the PH² dataset, DONet reaches a Dice score of 0.931, demonstrating strong generalization across datasets.
• Ablation studies confirm that both the dual objective architecture (DOA) and RCEM contribute significantly, with DOA improving Dice by 0.017 and RCEM by 0.019 over the baseline.
• The combination of DOA and RCEM yields the best performance, with Dice increasing to 0.881 on ISIC 2018 and 0.931 on PH², indicating synergistic effects.
• The RCEM with four time steps (dilation rates 1,2,4,8) achieves optimal performance, with diminishing returns beyond four steps.
• Visual ablation results show that both DOA and RCEM improve boundary adherence and reduce false positives, especially in challenging cases with small or irregular lesions.