[Paper Review] Small Organ Segmentation in Whole-body MRI using a Two-stage FCN and Weighting Schemes

Vanya V. Valindria, Ioannis Lavdas+5 more|arXiv (Cornell University)|Jul 30, 2018

Radiomics and Machine Learning in Medical Imaging15 references4 citations

TL;DR

This paper proposes a two-stage fully convolutional network (FCN) with class weighting and spatial priors to improve small organ segmentation in whole-body MRI. By combining weighted-FCN for coarse segmentation, auto-context for contextual refinement, and multi-atlas priors for localization, the method achieves a 0.560 DSC on small organs—significantly outperforming baseline FCN (0.244 DSC) and state-of-the-art methods.

ABSTRACT

Accurate and robust segmentation of small organs in whole-body MRI is difficult due to anatomical variation and class imbalance. Recent deep network based approaches have demonstrated promising performance on abdominal multi-organ segmentations. However, the performance on small organs is still suboptimal as these occupy only small regions of the whole-body volumes with unclear boundaries and variable shapes. A coarse-to-fine, hierarchical strategy is a common approach to alleviate this problem, however, this might miss useful contextual information. We propose a two-stage approach with weighting schemes based on auto-context and spatial atlas priors. Our experiments show that the proposed approach can boost the segmentation accuracy of multiple small organs in whole-body MRI scans.

Motivation & Objective

• Address the challenge of accurate small organ segmentation in whole-body MRI due to extreme class imbalance and weak boundaries.
• Overcome limitations of standard deep learning models that underperform on small organs due to low pixel occupancy and complex background.
• Improve segmentation accuracy by integrating coarse-to-fine architecture with contextual and anatomical priors.
• Evaluate the impact of class weighting, auto-context, and spatial priors on multi-organ segmentation performance in MRI.
• Extend beyond single-organ or CT-based methods to multiple small organs and bones in whole-body MRI.

Proposed method

• Apply a two-stage FCN: first stage uses a weighted-FCN to generate coarse segmentation and probability maps for all organs.
• Implement class weighting in the loss function to address class imbalance, giving higher loss contribution to underrepresented small organs.
• Use auto-context in the fine stage, where the coarse segmentation probability map is fed as an additional input to guide the fine network.
• Apply spatial priors from multi-atlas segmentation by multiplying the coarse probability map with atlas-based predictions to refine localization.
• Crop region-of-interest (ROI) for each organ using organ-specific bounding boxes derived from coarse predictions to focus on relevant anatomy.
• Optimize both coarse and fine networks jointly using end-to-end training with multi-task supervision.

Experimental results

Research questions

• RQ1 Can a two-stage FCN with class weighting significantly improve segmentation accuracy for small organs in whole-body MRI?
• RQ2 How does incorporating auto-context and spatial priors from multi-atlas segmentation affect fine-scale segmentation performance?
• RQ3 To what extent do weighting schemes mitigate class imbalance in small organ segmentation with less than 0.007% volume occupancy?
• RQ4 Does the combination of coarse-to-fine processing, auto-context, and spatial priors outperform state-of-the-art methods on multiple small organs?
• RQ5 How do different priors (e.g., multi-atlas vs. auto-context) impact segmentation of specific small organs like adrenal glands and clavicles?

Key findings

• The proposed two-stage weighted-FCN with auto-context and multi-atlas spatial priors achieved a mean DSC of 0.560 on small organs, a significant improvement over the baseline FCN (0.244 DSC).
• Class weighting alone increased small organ DSC from 0.244 to 0.483, demonstrating its effectiveness in mitigating class imbalance.
• The addition of multi-atlas priors improved segmentation for the adrenal gland (DSC: 0.203) and other small organs, though performance degraded slightly for clavicles due to low atlas accuracy.
• Auto-context integration reduced false positives and improved localization, especially for small and low-contrast organs like the adrenal glands.
• The method outperformed a state-of-the-art two-stage approach by approximately 50% in small organ DSC, confirming the benefit of joint optimization with contextual and spatial priors.
• The combination of weighting, auto-context, and spatial priors yielded the best overall performance, with the adrenal gland showing the most substantial relative improvement due to its minimal volume (0.0001% of total volume).