Star Shape Prior in Fully Convolutional Networks for Skin Lesion Segmentation

Zahra Mirikharaji, Ghassan Hamarneh

Abstract

Semantic segmentation is an important preliminary step towards automatic medical image interpretation. Recently deep convolutional neural networks have become the first choice for the task of pixel-wise class prediction. While incorporating prior knowledge about the structure of target objects has proven effective in traditional energy-based segmentation approaches, there has not been a clear way for encoding prior knowledge into deep learning frameworks. In this work, we propose a new loss term that encodes the star shape prior into the loss function of an end-to-end trainable fully convolutional network (FCN) framework. We penalize non-star shape segments in FCN prediction maps to guarantee a global structure in segmentation results. Our experiments demonstrate the advantage of regularizing FCN parameters by the star shape prior and our results on the ISBI 2017 skin segmentation challenge data set achieve the first rank in the segmentation task among 21 participating teams.

Introduction

Automated semantic segmentation of medical images, such as isolating skin lesions for melanoma detection, is a critical task in computer-aided diagnosis. Despite the success of Fully Convolutional Networks (FCNs), these models often struggle with global anatomical consistency. Because standard loss functions like binary cross-entropy evaluate pixels independently, the resulting masks can be fragmented or contain holes, failing to reflect the cohesive nature of biological structures.

Historically, traditional energy-based methods like active contours or graph cuts utilized shape priors to enforce geometric constraints. This paper bridges the gap between traditional geometric priors and modern deep learning by introducing a "star shape prior" directly into the FCN training process as a novel regularization term.

Methodology: The Star Shape Prior

The core contribution of this work is the star shape regularized loss ($L_{sh}$). An object is defined as star-shaped with respect to a center $c$ if every point $p$ within the object has a direct line of sight to $c$. Mathematically, for any point $p$ in the foreground, all points on the line segment $l_{pc}$ must also belong to the foreground.

Geometric definition of star shape (a) and examples of violations (b, c) where points lack a direct line of sight to the center.

During training, the network identifies the object's center from the ground truth mask. The $L_{sh}$ term penalizes instances where the network predicts a foreground pixel at $p$ but a background pixel at some point $q$ closer to the center on the same line $l_{pc}$. By incorporating this penalty into the loss function, the network learns to favor anatomically plausible, cohesive shapes. Crucially, this prior is only used during training; at inference time, the model performs a standard forward pass without requiring center-point knowledge or expensive post-processing like Conditional Random Fields (CRFs).

Experimental Results

The proposed method was evaluated on the ISIC 2017 Skin Lesion Analysis dataset, which contains 2,000 training and 600 test dermoscopy scans. Skin lesions are ideal candidates for this prior, as they generally form cohesive blobs; the authors verified that 99.86% of ground truth pixels in the dataset adhered to the star shape definition.

A representative dermoscopy scan of a skin lesion from the ISIC 2017 dataset.

The inclusion of the star shape prior yielded significant quantitative gains, improving the Jaccard index (Intersection over Union) by over 3% compared to baseline FCNs across different architectures. The best-performing model, a ResNet-DUC with star shape regularization, achieved a Jaccard index of 77.3% and a Dice score of 85.7%. This approach secured the first-place rank in the ISBI 2017 challenge, outperforming 20 other teams without the need for complex ensembling or multi-scale inputs. Qualitatively, the regularized masks were smoother and free of the fragmented "noise" typical of standard pixel-wise segmentation.

Conclusion

This research demonstrates that global geometric priors can be effectively integrated into deep learning frameworks via custom loss functions. By penalizing non-star-shaped predictions, the authors achieved state-of-the-art performance in skin lesion segmentation with zero inference-time overhead. While highly effective for lesions and tumors, future work will explore extending this framework to more complex 3D anatomical structures and irregular topologies where the star-shape assumption may need modification.

Summary updated

Abstract

Semantic segmentation is an important preliminary step towards automatic medical image interpretation. Recently deep convolutional neural networks have become the first choice for the task of pixel-wise class prediction. While incorporating prior knowledge about the structure of target objects has proven effective in traditional energy-based segmentation approaches, there has not been a clear way for encoding prior knowledge into deep learning frameworks. In this work, we propose a new loss term that encodes the star shape prior into the loss function of an end-to-end trainable fully convolutional network (FCN) framework. We penalize non-star shape segments in FCN prediction maps to guarantee a global structure in segmentation results. Our experiments demonstrate the advantage of regularizing FCN parameters by the star shape prior and our results on the ISBI 2017 skin segmentation challenge data set achieve the first rank in the segmentation task among $21$ participating teams.
BibTeX
@inproceedings{mirikharaji2018star,
  author = {Zahra Mirikharaji and Ghassan Hamarneh},
  title = {{Star Shape Prior in Fully Convolutional Networks for Skin Lesion Segmentation}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2018},
  doi = {10.1007/978-3-030-00937-3\_84},
}
Figure