ESTRO 2025 - Abstract Book

S3858

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Purpose/Objective: We used principal component analysis (PCA) to reduce the number of features and overcome multicollinearity of radiomic and deep neural network (DNN) features. Using head and neck cancer (HNC) PET/CT images to predict recurrence-free survival (RFS), we demonstrate that dimension reduction decreases noise and increases the performance of non-linear Cox proportional hazard models based on DNN and radiomics. Material/Methods: We analyzed the HNC PET/CT images from HECKTOR challenge [1]. After applying PCA on radiomic or DNN features, we selected principal components (PCs) to explain >90% of variance. Images were isotropically resampled to 2mm slice spacing, followed by training a 3D nnU-Net to segment lesions [2]. We extracted radiomics features from lesions using the PyRadiomics library, yielding a total amount of 424 features per sample, followed by normalization. Two DNN architectures were used: A) training a DenseNet [3] from scratch using PET/CT scans, B) fine-tuning a Med3D model [4] only with PET scans. In both, we used the CoxPHLoss proposed in [5], which is essentially the negative logarithm of the Cox likelihood:

A survival model trained with this loss function can be seen as a non-linear extension of the Cox proportional hazards model.

Results: Radiomic features exhibited substantial multicollinearity and thus training our survival model on them was not possible. Therefore, we trained the survival model with the CoxPHLoss using top 15 PCs, achieving a C-index of 0.667 (top score from the HECKTOR challenge was 0.682). We used five cross-validation (80/20 splits) for DNN models, which resulted in the test set C-index ranging [0.57, 0.63] for DenseNet and [0.56, 0.61] for Med3D. The best performing model with 0.63 C-index was used in our downstream analysis. With the top 5 PCs, our survival model further improved over the best performing end-to-end DNN to achieve a C-index of 0.645. We combined the radiomics and DNN features to train the unified Cox PH model, retaining the top 10 PCs. The Cox PH model trained on the combined feature set achieves a C-index of 0.659. Conclusion: PCA provides substantial dimension reduction which overcomes multicollinearity. Furthermore, using a small set of PCs is essentially performing noise reduction. The performance improvement after dimension reduction by PCA can be seen as reducing overfitting in a large number of potentially noisy features of DNN and radiomics. Lastly, we consider utilizing a small set of PCs to be more interpretable and computationally efficient. References: [1] Andrearczyk V, et al. Overview of the HECKTOR Challenge at MICCAI 2021: Automatic Head and Neck Tumor Segmentation and Outcome Prediction in PET/CT Images, in: Proceedings of HECKTOR 2021. LNCS (2022). [2] F. Isensee, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation, Nature Methods (2021). [3] G. Huang, F. Isensee, et al. Densely connected convolutional networks, in: Proceedings of the IEEE Conference on CVPR, (2017). [4] S. Chen, K. Ma, Y. Zheng, Med3D: Transfer learning for 3D medical image analysis, arXiv preprint arXiv:1904.00625 (2019). Keywords: Principal Components, survival, deep learning