Semi-supervised learning towards automated segmentation of PET images with limited annotations: application to lymphoma patients.

in Physical and engineering sciences in medicine by Fereshteh Yousefirizi, Isaac Shiri, Joo Hyun O, Ingrid Bloise, Patrick Martineau, Don Wilson, François Bénard, Laurie H Sehn, Kerry J Savage, Habib Zaidi, Carlos F Uribe, Arman Rahmim

TLDR

  • The study is about using a type of artificial intelligence called a convolutional neural network (CNN) to help doctors identify tumors in medical images. The study uses unlabeled data and focuses on PET images of diffuse large B-cell lymphoma (DLBCL) and primary mediastinal large B-cell lymphoma (PMBCL). The study compares different ways of training the CNN and found that using a combination of supervised and unsupervised learning approaches with a unified loss function was the best way to accurately identify tumors in the images. The study also found that a specific combination of supervision levels and loss functions was the best for automating medical imaging segmentation workflows. The study has important implications for the field of research and clinical practice, as it suggests that semi-supervised approaches can help doctors identify tumors more accurately and efficiently.

Abstract

Manual segmentation poses a time-consuming challenge for disease quantification, therapy evaluation, treatment planning, and outcome prediction. Convolutional neural networks (CNNs) hold promise in accurately identifying tumor locations and boundaries in PET scans. However, a major hurdle is the extensive amount of supervised and annotated data necessary for training. To overcome this limitation, this study explores semi-supervised approaches utilizing unlabeled data, specifically focusing on PET images of diffuse large B-cell lymphoma (DLBCL) and primary mediastinal large B-cell lymphoma (PMBCL) obtained from two centers. We considered 2-[F]FDG PET images of 292 patients PMBCL (n = 104) and DLBCL (n = 188) (n = 232 for training and validation, and n = 60 for external testing). We harnessed classical wisdom embedded in traditional segmentation methods, such as the fuzzy clustering loss function (FCM), to tailor the training strategy for a 3D U-Net model, incorporating both supervised and unsupervised learning approaches. Various supervision levels were explored, including fully supervised methods with labeled FCM and unified focal/Dice loss, unsupervised methods with robust FCM (RFCM) and Mumford-Shah (MS) loss, and semi-supervised methods combining FCM with supervised Dice loss (MS + Dice) or labeled FCM (RFCM + FCM). The unified loss function yielded higher Dice scores (0.73 ± 0.11; 95% CI 0.67-0.8) than Dice loss (p value < 0.01). Among the semi-supervised approaches, RFCM + αFCM (α = 0.3) showed the best performance, with Dice score of 0.68 ± 0.10 (95% CI 0.45-0.77), outperforming MS + αDice for any supervision level (any α) (p < 0.01). Another semi-supervised approach with MS + αDice (α = 0.2) achieved Dice score of 0.59 ± 0.09 (95% CI 0.44-0.76) surpassing other supervision levels (p < 0.01). Given the time-consuming nature of manual delineations and the inconsistencies they may introduce, semi-supervised approaches hold promise for automating medical imaging segmentation workflows.

Overview

  • The study explores the use of convolutional neural networks (CNNs) for accurately identifying tumor locations and boundaries in PET scans of diffuse large B-cell lymphoma (DLBCL) and primary mediastinal large B-cell lymphoma (PMBCL).
  • The study utilizes unlabeled data and focuses on PET images obtained from two centers, with 292 patients (104 PMBCL and 188 DLBCL) used for training and validation, and 60 patients for external testing. The study considers various supervision levels, including fully supervised, unsupervised, and semi-supervised methods, and incorporates both supervised and unsupervised learning approaches. The primary objective is to identify the best semi-supervised approach for automating medical imaging segmentation workflows.

Comparative Analysis & Findings

  • The study compares the outcomes observed under different experimental conditions, specifically focusing on various supervision levels for training the 3D U-Net model. The study identifies significant differences in the results between these conditions, with the unified loss function yielding higher Dice scores than Dice loss (p value < 0.01). Among the semi-supervised approaches, RFCM + αFCM (α = 0.3) showed the best performance, with a Dice score of 0.68 ± 0.10 (95% CI 0.45-0.77), outperforming MS + αDice for any supervision level (any α) (p < 0.01). Another semi-supervised approach with MS + αDice (α = 0.2) achieved a Dice score of 0.59 ± 0.09 (95% CI 0.44-0.76) surpassing other supervision levels (p < 0.01).
  • The key findings of the study suggest that semi-supervised approaches hold promise for automating medical imaging segmentation workflows, with RFCM + αFCM (α = 0.3) showing the best performance among the semi-supervised approaches. The study also identifies the unified loss function as a promising approach for improving the accuracy of medical imaging segmentation.

Implications and Future Directions

  • The study's findings have significant implications for the field of research and clinical practice, as they suggest that semi-supervised approaches can automate medical imaging segmentation workflows, reducing the time-consuming nature of manual delineations and the inconsistencies they may introduce. The study identifies several limitations, including the need for further validation on larger datasets and the potential for overfitting. Future research directions could explore the use of transfer learning for improving the performance of semi-supervised approaches, as well as the integration of additional modalities, such as CT scans, for improving the accuracy of tumor segmentation.
Read Full Article