[Paper Review] Learnable Hollow Kernels for Anatomical Segmentation.

Elizaveta Lazareva, Oleg Y. Rogov+3 more|arXiv (Cornell University)|Jul 9, 2020

Advanced Neural Network Applications20 references4 citations

TL;DR

This paper proposes learnable hollow kernels that mimic the shape and structural complexity of hollow organs like the bladder to improve anatomical segmentation in medical imaging. By integrating these adaptive kernels into U-Net-like networks with spatio-temporal modeling (e.g., Bi-LSTM), the method achieves state-of-the-art performance, attaining mean Dice scores of 0.936, 0.736, and 0.712 for inner wall, outer wall, and tumor regions, respectively.

ABSTRACT

Segmentation of certain hollow organs, such as the bladder, is especially hard to automate due to their complex geometry, vague intensity gradients in the soft tissues, and a tedious manual process of the data annotation routine. Yet, accurate localization of the walls and the cancer regions in the radiologic images of such organs is an essential step in oncology. To address this issue, we propose a new class of hollow kernels that learn to 'mimic' the contours of the segmented organ, effectively replicating its shape and structural complexity. We train a series of the U-Net-like neural networks using the proposed kernels and demonstrate the superiority of the idea in various spatio-temporal convolution scenarios. Specifically, the dilated hollow-kernel architecture outperforms state-of-the-art spatial segmentation models, whereas the addition of temporal blocks with, e.g., Bi-LSTM, establishes a new multi-class baseline for the bladder segmentation challenge. Our spatio-temporal model based on the hollow kernels reaches the mean dice scores of 0.936, 0.736, and 0.712 for the bladder's inner wall, the outer wall, and the tumor regions, respectively. The results pave the way towards other domain-specific deep learning applications where the shape of the segmented object could be used to form a proper convolution kernel for boosting the segmentation outcome.

Motivation & Objective

• To address the challenge of automating segmentation in hollow organs with complex geometry and weak intensity gradients, such as the bladder.
• To reduce the burden of tedious manual annotation by developing a shape-aware deep learning approach.
• To improve segmentation accuracy for critical structures—especially inner and outer walls and tumor regions—by leveraging the organ's intrinsic shape in the convolutional kernel design.
• To establish a new multi-class baseline for bladder segmentation using spatio-temporal modeling with learnable hollow kernels.

Proposed method

• Propose a novel class of learnable hollow kernels that dynamically adapt to the shape and structural complexity of target organs like the bladder.
• Integrate these hollow kernels into U-Net-like architectures to enhance spatial feature learning in 3D medical images.
• Apply dilated convolutions with hollow kernels to capture long-range dependencies while preserving fine anatomical details.
• Incorporate temporal modeling blocks, such as Bi-LSTM, into the network to exploit temporal consistency in dynamic MRI sequences.
• Train the end-to-end network using supervised learning with ground-truth segmentation masks to optimize kernel parameters and segmentation heads simultaneously.
• Use a differentiable kernel generation mechanism that allows the network to learn the kernel shape from the data, effectively mimicking organ contours.

Experimental results

Research questions

• RQ1 Can learnable hollow kernels that mimic organ shape improve segmentation accuracy in complex hollow organs like the bladder?
• RQ2 How does the integration of dilated hollow kernels compare to standard convolutional layers in spatial-only segmentation tasks?
• RQ3 To what extent does adding temporal modeling with Bi-LSTM layers enhance performance in dynamic MRI sequences for bladder segmentation?
• RQ4 Can the proposed hollow kernel architecture generalize to multi-class segmentation of inner wall, outer wall, and tumor regions?
• RQ5 What is the quantitative performance gain of the spatio-temporal hollow kernel model over existing state-of-the-art methods?

Key findings

• The dilated hollow-kernel architecture significantly outperforms state-of-the-art spatial segmentation models in bladder segmentation.
• The spatio-temporal model with Bi-LSTM temporal blocks achieves a mean Dice score of 0.936 for the bladder's inner wall.
• The model reaches a mean Dice score of 0.736 for the outer wall, demonstrating strong performance on the more challenging outer wall segmentation.
• The model achieves a mean Dice score of 0.712 for tumor regions, indicating effective detection of pathological structures.
• The proposed method establishes a new multi-class baseline for the bladder segmentation challenge, surpassing existing approaches in both spatial and temporal modeling scenarios.
• The learnable hollow kernels effectively capture the complex geometry of hollow organs, leading to improved segmentation fidelity and robustness.