Multimodal PET/CT-based PD-L1 status prediction in lung cancer via semi-supervised and unsupervised deep learning.
Authors
Affiliations (8)
Affiliations (8)
- LaTIM, UMR 1101, Inserm, University of Brest, Brest, France. [email protected].
- LaTIM, UMR 1101, Inserm, University of Brest, Brest, France.
- Nuclear Medicine, University of Poitiers, Poitiers, France.
- Radiation Oncology Department, University Hospital, Brest, France.
- Division of Nuclear Medicine and Oncological Imaging, University Hospital of Liège, Liège, Belgium.
- Division of Pulmonology, CHU Liège, Liège, Belgium.
- Department of Radiotherapy Oncology, University Hospital of Liège, Liège, Belgium.
- GIGA-CRC In Vivo Imaging, University of Liège, Liège, Belgium.
Abstract
Deep learning (DL) techniques have been applied in lung cancer screening, assessing drug effectiveness, and enhancing prognosis prediction. Within this context, the combination of 18FDG PET/CT images with DL has demonstrated promising results, particularly in predicting programmed death ligand-1 (PD-L1) expression in lung cancer, improving overall prediction accuracy and offering a viable non-invasive complementary imaging biomarker to support clinical decision-making and patient stratification. An effective way to improve the performance of deep neural networks in most tasks is to increase the quantity of labeled data and the quality of labels. However, in medical imaging, high-quality annotations and large datasets are both challenging to obtain due to the need for expert knowledge and tedious procedures including regulatory obstacles. In this context, we propose a semi-supervised and unsupervised deep neural networks (USSLNet) using early fusion multi-modal PET/CT images within the context of predicting PD-L1 expression. By alternately running two tasks, label information is propagated to the unlabeled data, enabling the model to extract semantic information and mitigating the risk of overfitting to limited labeled data. Model performance was evaluated using the area under the receiver operating characteristic curves (AUCs) and 95% confidence intervals (CIs). Compared with current methods, our framework demonstrates improved robustness, reducing the impact of outliers and yielding superior performance in PD-L1 status classification. Moreover, the framework consistently outperformed current approaches when utilizing various types of unlabeled PET/CT images. These findings highlight the effectiveness of our approach in predicting PD-L1 expression through the use of limited in size and partially annotated multi-modal PET/CT datasets.