Ground Truth base Deep Learning Model Initialization for Medical Image Segmentation

Zijin Yang, Nazmus Sakib, Chang Liu, Andreas Maier

Keywords: deep learning, kernel initialization, medical imaging, segmentation

TL;DR: a kernel initialization method using ground truth information for optimal convergence

Abstract:

Deep learning-based approaches for medical image segmentation have demonstrated considerable promise; however, their clinical translation remains challenging due to limited annotated datasets and the suboptimal performance of standard weight initialization strategies when applied to volumetric medical imaging data. To address these limitations, we propose a novel data-driven initialization framework, termed Principal Component Analysis-based Reference Patient initialization (PR-PCA), which leverages prior anatomical knowledge extracted from a reference patient to customize convolutional kernel weights. Our method employs Principal Component Analysis (PCA) to decompose structural information from the reference patient, which is then systematically propagated to initialize kernels across network depths, thereby encoding domain-specific anatomical priors into the network architecture.

We systematically evaluated the proposed initialization strategy against conventional Xavier and Kaiming initialization methods using U-Net and Residual U-Net architectures on two distinct medical imaging datasets. Experiments were conducted with varying kernel configurations (3 × 3 × 3 and 7 × 7 × 7) across multiple training iterations. Quantitative assessment was performed using the Dice similarity coefficient as the primary evaluation metric.

Our results demonstrate that the PR-PCA initialization consistently outperforms baseline methods across both kernel configurations. For the 3 × 3 × 3 kernel architecture trained over 1000 epochs, PR-PCA achieved a Dice score of 0.79, representing improvements of 3.9\% and 31\% over Kaiming (0.76) and Xavier (0.48) initialization, respectively. Similarly, for the 7 × 7 × 7 kernel configuration trained over 500 epochs, PR-PCA achieved a Dice score of 0.79, compared to 0.76 for both Xavier and Kaiming methods. These findings validate that incorporating anatomical priors through PCA-based initialization significantly enhances segmentation performance, particularly in scenarios with limited training data. This approach provides a principled framework for integrating domain knowledge into deep learning models for medical image analysis.

Primary Subject Area: Segmentation

Secondary Subject Area: Transfer Learning and Domain Adaptation

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 350