Click here to


Are you sure ?

Yes, do it No, cancel

Investigating Radiomics to Predict Positive Lymph Nodes in Oral Cavity Squamous Cell Carcinoma (OSCC)

A Traverso12*, A Hosni-Abdalaty2 , M Hasan2 , J Kim2 , J Ringash2 , J Cho2 , S Bratman2 , A Bailey2 , J Waldron9 , M Welch2 , J Irish3 , B O'Sullivan2 , J De Almeida3 , M Giuliani2 , D Chepeha2 , D Goldstein2 , D Jaffray2 , L Wee1 , A Dekker1 , A Hope2 , (1) MAASTRO Clinic, Maastricht, The Netherlands ,(2) Princess Margaret Cancer Centre, Toronto, Canada (3) University Health Network, Toronto, Canada


(Tuesday, 7/16/2019) 7:30 AM - 9:30 AM

Room: Stars at Night Ballroom 2-3

Purpose: Radiomics features may provide additional information to support predictive models of clinical outcomes in a non-invasive manner. In this preliminary study, we combined radiomics with state-of-the-art machine learning (ML) algorithms for predicting positive lymph nodes in OSCC patients.

Methods: 134 patients with both biopsy proven OSCC and pre-operative diagnostic MRI who eventually underwent curative surgical extirpation including neck dissection with or without adjuvant radiotherapy were identified from our database of OSCC patient outcomes. The primary oral cavity gross tumour volume (GTV) was delineated in a blinded fashion on the pre-operative T2-weighted MRI without information on the results of subsequent neck dissection by a single physician. Different radiomics features classes (first order statistics, shape metrics and textural analysis) were extracted from the delineated GTV using PyRadiomics. Six different unsupervised ML methods were combined with three different classifiers to build the strongest radiomic signature predicting the presence/absence of positive lymph nodes in the patients after neck dissection. Training and validation were repeated 100 times using stratified subsampled portions of the data, 80 and 20% respectively, without replacement. The mode prediction for each patient across the 100 subsampled fittings were used to calculate the Area under the curve (AUC) for all the possible combinations.

Results: The best radiomic signature had an AUC of 0.83 on the training dataset and 0.81 on the validation. The best clinical model including tumour staging, localization of the tumour (tongue vs non-tongue), tumour thickness, gender and sex had an AUC of 0.65 on the training dataset and 0.60 on the validation dataset.

Conclusion: We proposed a method based on radiomics and ML to build a signature for predicting positive lymph nodes in OSCC patients. Results show the power of radiomics and ML applied to radiation oncology but need further validation on external datasets to verify their generalizability.


Modeling, MR


IM/TH- Image Analysis (Single modality or Multi-modality): Imaging biomarkers and radiomics

Contact Email