[Paper Review] Semi-supervised learning towards automated segmentation of PET images with limited annotations: Application to lymphoma patients

Fereshteh Yousefirizi, Isaac Shiri+10 more|arXiv (Cornell University)|Dec 19, 2022

Medical Imaging Techniques and Applications4 citations

TL;DR

This study proposes a semi-supervised 3D U-Net framework using fuzzy c-means (FCM) and Mumford-Shah (MS) losses to automate PET image segmentation in lymphoma patients with minimal expert annotations. The best-performing method, RFCM+alpha*LFCM with alpha=0.3, achieved a Dice score of 0.69 ± 0.03, outperforming both supervised and other semi-supervised approaches, demonstrating significant potential for reducing manual segmentation effort in clinical workflows.

ABSTRACT

The time-consuming task of manual segmentation challenges routine systematic quantification of disease burden. Convolutional neural networks (CNNs) hold significant promise to reliably identify locations and boundaries of tumors from PET scans. We aimed to leverage the need for annotated data via semi-supervised approaches, with application to PET images of diffuse large B-cell lymphoma (DLBCL) and primary mediastinal large B-cell lymphoma (PMBCL). We analyzed 18F-FDG PET images of 292 patients with PMBCL (n=104) and DLBCL (n=188) (n=232 for training and validation, and n=60 for external testing). We employed FCM and MS losses for training a 3D U-Net with different levels of supervision: i) fully supervised methods with labeled FCM (LFCM) as well as Unified focal and Dice loss functions, ii) unsupervised methods with Robust FCM (RFCM) and Mumford-Shah (MS) loss functions, and iii) Semi-supervised methods based on FCM (RFCM+LFCM), as well as MS loss in combination with supervised Dice loss (MS+Dice). Unified loss function yielded higher Dice score (mean +/- standard deviation (SD)) (0.73 +/- 0.03; 95% CI, 0.67-0.8) compared to Dice loss (p-value<0.01). Semi-supervised (RFCM+alpha*LFCM) with alpha=0.3 showed the best performance, with a Dice score of 0.69 +/- 0.03 (95% CI, 0.45-0.77) outperforming (MS+alpha*Dice) for any supervision level (any alpha) (p<0.01). The best performer among (MS+alpha*Dice) semi-supervised approaches with alpha=0.2 showed a Dice score of 0.60 +/- 0.08 (95% CI, 0.44-0.76) compared to another supervision level in this semi-supervised approach (p<0.01). Semi-supervised learning via FCM loss (RFCM+alpha*LFCM) showed improved performance compared to supervised approaches. Considering the time-consuming nature of expert manual delineations and intra-observer variabilities, semi-supervised approaches have significant potential for automated segmentation workflows.

Motivation & Objective

• To address the time-consuming and variable nature of manual tumor segmentation in 18F-FDG PET scans for lymphoma patients.
• To reduce reliance on extensive expert-annotated data by leveraging semi-supervised learning in medical image segmentation.
• To evaluate and compare the performance of different loss functions—FCM, MS, and unified focal/Dice—under varying supervision levels.
• To identify the optimal semi-supervised configuration that maximizes segmentation accuracy with limited labeled data.
• To develop a robust, automated segmentation pipeline applicable to diffuse large B-cell lymphoma (DLBCL) and primary mediastinal B-cell lymphoma (PMBCL).

Proposed method

• A 3D U-Net architecture was trained using a combination of labeled and pseudo-labeled data to enable semi-supervised learning.
• Fuzzy C-Means (FCM) and Robust FCM (RFCM) were used as unsupervised loss functions to guide clustering of image features.
• The Mumford-Shah (MS) loss function was applied to enforce piecewise smoothness in segmentation boundaries.
• Unified loss functions combining focal and Dice losses improved training stability and segmentation performance.
• Semi-supervised training combined RFCM with supervised Dice loss (RFCM + alpha*LFCM) and MS with Dice (MS + alpha*Dice), with hyperparameter alpha controlling supervision balance.
• Model performance was evaluated on 60 external test cases from 292 patients with DLBCL and PMBCL, using Dice score as the primary metric.

Experimental results

Research questions

• RQ1 Can semi-supervised learning with FCM and MS losses significantly improve PET image segmentation accuracy when labeled data is limited?
• RQ2 How does the integration of robust FCM (RFCM) with supervised loss functions compare to purely supervised or unsupervised approaches?
• RQ3 What is the optimal balance between supervised and unsupervised loss components (via hyperparameter alpha) for maximum segmentation performance?
• RQ4 Does the unified focal and Dice loss function outperform standard Dice loss in semi-supervised settings?
• RQ5 Can the proposed method reduce inter-observer variability and manual delineation time in clinical lymphoma PET imaging?

Key findings

• The unified focal and Dice loss function yielded a higher Dice score (0.73 ± 0.03) than standard Dice loss (p < 0.01), indicating improved training stability and performance.
• The semi-supervised approach using RFCM + 0.3*LFCM achieved the best overall performance with a Dice score of 0.69 ± 0.03 (95% CI: 0.45–0.77), significantly outperforming all other configurations (p < 0.01).
• The MS + 0.2*Dice semi-supervised method achieved a Dice score of 0.60 ± 0.08 (95% CI: 0.44–0.76), which was inferior to RFCM+LFCM across all alpha values (p < 0.01).
• Semi-supervised learning with RFCM and labeled data (RFCM+alpha*LFCM) outperformed fully supervised methods, demonstrating the value of pseudo-labeling in low-data regimes.
• The proposed method significantly reduces dependency on expert-annotated data while maintaining high segmentation accuracy, suggesting strong clinical applicability.
• The results indicate that combining robust unsupervised clustering (RFCM) with supervised loss functions is a promising strategy for automated tumor segmentation in oncology PET imaging.