Abstract
Accurate lesion segmentation is a cornerstone of computer-assisted diagnosis, yet it remains challenging due to low contrast, irregular boundaries, and high variability in lesion size. This paper presents Deep Active Lesion Segmentation (DALS), a hybrid framework that integrates the high-level feature extraction of Convolutional Neural Networks (CNNs) with the precise boundary-delineation of classical Active Contour Models (ACMs). By utilizing a CNN to automatically initialize and provide per-pixel tuning parameters for a level-set evolution, DALS achieves state-of-the-art accuracy across multiple organs and imaging modalities.
Introduction
The segmentation of medical lesions is critical for clinical decision-making. Current methodologies typically fall into two categories: Active Contour Models (ACMs) and Deep Learning (CNNs). ACMs use mathematical energy functionals to evolve curves toward boundaries but often require manual initialization and are prone to local minima. Conversely, CNNs like U-Net excel at localization but frequently produce imprecise or "blurry" boundaries. DALS bridges this gap by creating a fully automated pipeline where deep learning "drives" the classical mathematics of active contours.
Methods
The DALS framework consists of two primary components: a CNN backbone and a parameterized Level-Set ACM.
The CNN backbone utilizes an encoder-decoder architecture with Dense Blocks and a Multiscale Dilation Block to capture both local textures and global context simultaneously. The workflow proceeds as follows:
- The CNN processes an input image to generate a probability map ($Y_{prob}$).
- A Transformer module converts this map into a Signed Distance Map (SDM), providing an automated initialization for the contour.
- The probability map is used to calculate per-pixel parameter functions ($\lambda_1(x,y)$ and $\lambda_2(x,y)$), which act as spatially-varying weights for contour evolution.
- The contour evolves by solving a Partial Differential Equation (PDE) until it settles on the lesion boundary.
Experimental Setup
The model was evaluated on a new Multiorgan Lesion Segmentation (MLS) dataset, comprising:
- Brain: 369 MRI perfusion scans.
- Liver: 112 CT scans and 164 MRI scans.
- Lung: 87 CT scans.
This dataset is particularly challenging due to high variance in lesion radius and varying levels of global contrast and heterogeneity.
Results
DALS was compared against standard U-Net, the standalone CNN backbone, and manually-initialized level-sets.
DALS achieved superior accuracy, with the highest Dice Scores across all categories: Brain MR (0.888), Lung CT (0.869), Liver MR (0.894), and Liver CT (0.846). Notably, DALS significantly outperformed U-Net in Hausdorff Distance, demonstrating that the ACM refinement produces much sharper and more anatomically correct edges. Wilcoxon paired tests confirmed these improvements were statistically significant ($p < 0.001$). Despite the added complexity of solving PDEs, total inference time remains efficient at approximately 1.5 seconds per image.
Discussion and Conclusion
The key innovation of DALS lies in its transition from global scalar parameters to per-pixel parameter maps, allowing for flexible contour evolution in uncertain areas. By incorporating geometric constraints, the ACM acts as a regularizer, enabling high performance even with limited training data. While currently implemented for 2D slices, future work will explore 3D volumetric segmentation and further optimization of the parameter mapping functions. DALS represents a significant step toward robust, fully automated lesion segmentation in clinical workflows.
Summary updated