ESTRO 2024 - Abstract Book

S4528

Physics - Machine learning models and clinical applications

ESTRO 2024

[5] Chen J, Guo H, Yi K, Li B, and Elhoseiny M. VisualGPT: Data-efficient adaptation of pretrained language models for image captioning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18009– 18019, 2022. [6] Haidar A, Field M, Batumalai V, Cloak K, Al Mouiee D, Chlap P, Huang X, Chin V, Aly F, Carolan M, Sykes J, Vinod SK, G Delaney, and L Holloway. Standardising breast radiotherapy structure naming conventions: A machine learning approach. Cancers, 15(3), 2023.

2234

Digital Poster

Differentiation of pseudo- and true-progression in glioma: a self-supervised deep learning approach

Ahmed Gomaa, Yixing Huang, Rainer Fietkau, Christoph Bert, Udo S Gaipl, Benjamin Frey, Florian Putz

University Hospital Erlangen, Department of Radiation Oncology, Erlangen, Germany

Purpose/Objective:

The accurate differentiation between pseudo-progression (PsP) and true-progression (TP) in radiation oncology, particularly in the context of High-Grade Glioma (HGG), is of paramount importance for timely clinical decision making. Pseudo-progression, a transient radiological phenomenon after chemoradiation, and true-progression, indicative of disease progression, often exhibit overlapping characteristics, posing a significant challenge in radiation oncology practice. In this study, we aim to develop and evaluate a deep learning model capable of distinguishing between these two distinct states to enhance the accuracy of diagnosis and treatment planning.

Material/Methods:

The proposed framework leverages the availability of rich unlabeled multi-parametric MRI data of glioma cases. To achieve this, we employ a Vision Transformer (ViT) as an encoder, which is responsible for extracting concise, clinically relevant representation embeddings from the initially complex MRI data in a self-supervised manner [1]. We chose the ViT because it excels at capturing both global and local context from high-dimensional MRI data via its attention mechanism. This extraction process involves two essential tasks: context restoration, which helps the encoder understand the structural and anatomical context of brain regions, and contrastive learning, which improves the compactness and separability of these representations. The self-supervised approach is crucial for mitigating overfitting, ensuring the model's stability despite the limited data availability. Subsequently, we use the ViT to encode the MR volumes of glioma cases with known progression status, generating discriminative compact representation vectors. These vectors are then used to train the parameters of a fully connected network for the final progression status classification. In the self-supervised (upstream) task, we worked with the open BraTS 2021 dataset encompassing 1250 glioma cases [2-4], while the progression classification task incorporated 40 cases for training, distributed as [23 TP, 4 PsP, and 13 Mixed Response], and 20 cases for testing, distributed as [11 TP, 4 PsP, and 5 Mixed Response] from the open UPenn-