Weakly-supervised Cerebrovascular Segmentation Network with Shape Prior and Model Indicator

Abstract

Labeling cerebral vessels requires domain knowledge in neurology and could be extremely laborious, and there is a scarcity of public annotated cerebrovascular datasets. Traditional machine learning or statistical models could yield decent results on thick vessels with high contrast while having poor performance on those regions of low contrast. In our work, we employ a statistic model as noisy labels and propose a Transformer-based architecture which utilizes Hessian shape prior as soft supervision. It enhances the learning ability of the network to tubular structures, so that the model can make more accurate predictions on refined cerebrovascular segmentation. Furthermore, to combat the overfitting towards noisy labels as model training, we introduce an effective label extension strategy that only calls for a few manual strokes on one sample. These supplementary labels are not used for supervision but only as an indicator to tell where the model keeps the most generalization capability, so as to further guide the model selection in validation. Our experiments are carried out on a public TOF-MRA dataset from MIDAS data platform, and the results demonstrate that our method shows superior performance on cerebrovascular segmentation which achieves Dice of 0.831±0.040 in the dataset.

A review on AI-based medical image computing in head and neck surgery

“…Some researchers have conducted some studies on the head & neck anatomy. For example, Wu et al (2022) proposed a weakly supervised transformer network based on shape prior and model indicator, which can better segment low-contrast regions. Chen et al (2022a) introduced the transformer into a semi-supervised network for segmenting cerebrovascular on magnetic resonance angiography, which can effectively reduce the under-segmentation.…”

Understanding the brain with attention: A survey of transformers in brain sciences

“…It shows that the most classical optimization scheme is the combination of CE loss and Adam, which has been widely proven to be conducive to fast convergence [222,223] . When samples are not evenly distributed (i.e., cerebrovascular and brain tumor), the weight CE loss is a classical scheme to balance target constraints and has developed Cosine similarity [201] and Hessian soft weights [148] . In a segmentation scenario, the L2 and Dice regularization are often combined with CE loss to supplement empirical priors and reduce overfitting due to excessive predicted voxels.…”

“…[21,146] In Transformer-based models, UNETR is an important baseline for cerebrovascular segmentation. For example, Chen et al [147] and Wu et al [148] successfully performed cerebrovascular segmentation using UNETR in timeof-flight Magnetic Resonance (MR) angiography (TOF-MRA). Chen et al [147] further proposed a TRansformer with Semantic Fusion (TRSF-Net) for cerebrovascular segmentation.…”

“…For example, Chen et al [147] . and Wu et al [148] . successfully performed cerebrovascular segmentation using UNETR in time‐of‐flight Magnetic Resonance (MR) angiography (TOF‐MRA).…”

“…[222,223] When samples are not evenly distributed (i.e., cerebrovascular and brain tumor), the weight CE loss is a classical scheme to balance target constraints and has developed Cosine similarity [201] and Hessian soft weights. [148] In a segmentation scenario, the L2 and Dice regularization are often combined with CE loss to supplement empirical priors and reduce overfitting due to excessive predicted voxels. Moreover, some researches have proposed proprietary regularization, such as the FOG, [157] topological, [158] and Structural SIMilarity (SSIM) loss, [143] to be adapted to specific tasks.…”