Artificial intelligence classification of spatial CD8 + T-cell distribution on computed tomography in oropharyngeal cancer.
Authors
Affiliations (7)
Affiliations (7)
- Radiation Physics, Dept. of Hematology, Oncology & Radiation Physics, Skåne University Hospital, 222 42 Lund, Sweden; Dept. of Clinical Sciences, Lund University, 221 84 Lund, Sweden; Dept. of Hematology, Oncology & Radiation Physics, Skåne University Hospital, 222 42 Lund, Sweden.
- Dept. of Clinical Sciences, Lund University, 221 84 Lund, Sweden; Dept. of ORL, Head & Neck Surgery, Skåne University Hospital, 221 85 Lund, Sweden; Dept. of Hematology, Oncology & Radiation Physics, Skåne University Hospital, 222 42 Lund, Sweden.
- Dept. of Immunotechnology, Lund University, 223 81 Lund, Sweden.
- Centre for Mathematical Sciences, Lund University, 221 00 Lund, Sweden; Dept. of Hematology, Oncology & Radiation Physics, Skåne University Hospital, 222 42 Lund, Sweden.
- Radiation Physics, Dept. of Hematology, Oncology & Radiation Physics, Skåne University Hospital, 222 42 Lund, Sweden; Dept. of Hematology, Oncology & Radiation Physics, Skåne University Hospital, 222 42 Lund, Sweden.
- Radiation Physics, Dept. of Hematology, Oncology & Radiation Physics, Skåne University Hospital, 222 42 Lund, Sweden; Dept. of Translational Medicine, Medical Radiation Physics, Lund University, Malmö, Sweden; Dept. of Hematology, Oncology & Radiation Physics, Skåne University Hospital, 222 42 Lund, Sweden.
- Dept. of Immunotechnology, Lund University, 223 81 Lund, Sweden; Dept. of Hematology, Oncology & Radiation Physics, Skåne University Hospital, 222 42 Lund, Sweden.
Abstract
CD8 + T-cell presence and distribution in oropharyngeal cancer (OPC) reflects specific immune phenotypes and are linked to prognosis. Image processing methods such as convolutional neural networks (CNN) may enhance treatment decisions. From standard computed tomography (CT) images, this study evaluates radiomics in combination with machine learning models and CNN for prediction of spatially resolved CD8 + T-cell immune profiles and survival outcomes. Pre-treatment CT images were analysed using and comparing radiomics and CNNs to predict stroma-to-cancer cell islet CD8 + T-cell infiltration ratios (SIR) and immune phenotypes as determined from spatially resolved tissue-array. Open data with genomic and image information were used for external validation. Further, we predicted survival outcomes stratified with our model in OPC patients recruited to the ARTSCAN III randomised clinical trial, which compared cetuximab with cisplatin concurrent with radiotherapy. Our CNN model achieved an area under the curve (AUC) of 0.75 (range 0.71-0.81) in predicting high versus low SIR, performing better than radiomics. When stratifying the ARTSCAN III patients according to the CNN-derived immune scores, patients with an inflamed phenotype had improved local control (HR 0.03, 95% CI 0-0.23, p < 0.0001), progression-free survival (PFS) (HR 0.36, 95%CI 0.18-0.71, p = 0.003), and overall survival (HR 0.35, 95%CI 0.14-0.85, p = 0.02) in univariable and multivariable analyses. When stratifying the ARTSCAN III patients by type of chemotherapy, high immune score classification corresponded to better PFS in cetuximab-treated patients. Our findings indicate the potential of CNNs to predict CD8 + T-cell infiltration in OPC, offering a promising non-invasive tool for treatment stratification and personalised therapy.