Deep Learning-Based Outcome Prediction Using Harmonized CT Images for Head and Neck Cancer Patients: A Novel Approach with Attention and DeepHitSingle - Université d'Angers Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

Deep Learning-Based Outcome Prediction Using Harmonized CT Images for Head and Neck Cancer Patients: A Novel Approach with Attention and DeepHitSingle

Nassib Abdallah
  • Fonction : Auteur
Jean-Marion Marion
  • Fonction : Auteur
C. Tauber
  • Fonction : Auteur
Thomas Carlier
  • Fonction : Auteur
Mathieu Hatt
  • Fonction : Auteur

Résumé

This study presents a novel approach for predicting progression-free survival (PFS) in head and neck cancer patients, emphasizing on the importance of image harmonization and missing values imputation. We used the HECKTOR 2021 Challenge dataset containing 325 PET/CT scans (224 patients for training, 101 for testing) from six clinical centers, along with clinical data and contours of tumors. Previous studies achieved concordance index (C-index) scores between 0.61 and 0.72, but none explicitly harmonized multicentric images. Our two-module pipeline outperformed these previous results. The first module pre-processes the CT component of PET/CT images, addressing inter-scanner variability through a deep learning-based image harmonization model using a Multi-Scale Convolutional Neural Network (MCNN) with channel-wise and spatial attention mechanisms. The harmonized CT images were combined with clinical variables for which missing values were imputed using Multiple Correspondence Analysis (MCA) into our second module for PFS prediction relying on a MultiLayer Perceptron (MLP) NN integrated into a DeepHitSingle-based model. Our approach achieved a C-index of 0.75 using raw clinical features. With MCA, the C-index increased to 0.80. Ablation studies confirmed the effectiveness of MCNN for image harmonization and the DeepHit survival model. The addition of a fusion layer, combining channel-wise and spatial attention layers, further improved the C-index to 0.86 ± 0.035. Future work will focus on adding PET image harmonization to investigate the added value of combined PET/CT and evaluating our approach on the 2022 dataset (880 patients across 9 centers).
Fichier non déposé

Dates et versions

hal-04606873 , version 1 (10-06-2024)

Identifiants

Citer

Nassib Abdallah, Jean-Marion Marion, C. Tauber, Thomas Carlier, Pierre Chauvet, et al.. Deep Learning-Based Outcome Prediction Using Harmonized CT Images for Head and Neck Cancer Patients: A Novel Approach with Attention and DeepHitSingle. 2023 IEEE Nuclear Science Symposium, Medical Imaging Conference and International Symposium on Room-Temperature Semiconductor Detectors (NSS MIC RTSD), Nov 2023, Vancouver, France. pp.1-2, ⟨10.1109/NSSMICRTSD49126.2023.10338142⟩. ⟨hal-04606873⟩
7 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Mastodon Facebook X LinkedIn More