ESTRO 2022

Session Item

Tumour radiobiology

Session Code: 8008

Session Type: Poster (digital)

Track: Radiobiology

Regression and deep learning for transcriptome-based HPV-status prediction in head and neck cancer

Kristian Unger, Germany

Presentation Number: PO-1824

Abstract

Abstract Title:

Regression and deep learning for transcriptome-based HPV-status prediction in head and neck cancer

Authors:

Kristian Unger^1,2,3, Elia Lombardo^3,6, Julia Hess^4,5,3, Christopher Kurz^3,6, Marco Riboldi⁶, Sebastian Marschner^3,2, Philipp Baumeister^7,2, Kirsten Lauber^3,2, Ulrike Pflugradt^3,2, Axel Walch^8,2, Martin Canis^7,2, Frederick Klauschen^9,10, Horst Zitzelsberger^11,2,3, Claus Belka^3,2,10, Guillaume Landry^3,12

¹Helmholtz Zentrum München, German Research Center for Environmental Health GmbH, Radiation Cytogenetics, Neuherberg, Germany; ²Helmholtz Zentrum München, German Research Center for Environmental Health GmbH, Clinical Cooperation Group "Personalized Radiotherapy in Head and Neck Cancer", Neuherberg, Germany; ³University Hospital, LMU Munich, Department of Radiation Oncology, München, Germany; ⁴Helmholtz Zentrum München, German Research Center for Environmental Health GmbH, German Research Center for Environmental Health GmbH, Neuherberg, Germany; ⁵ Helmholtz Zentrum München, German Research Center for Environmental Health GmbH, Clinical Cooperation Group "Personalized Radiotherapy in Head and Neck Cancer", Neuherberg, Germany; ⁶Faculty of Physics, Ludwig-Maximilians-Universität München, Department of Medical Physics, Garching, Germany; ⁷University Hospital, LMU Munich, Department of Otorhinolaryngology, München, Germany; ⁸Helmholtz Zentrum München, German Research Center for Environmental Health GmbH, Research Unit Analytical Pathology, Neuherberg, Germany; ⁹Faculty of Medicine, Ludwig-Maximilians-University of Munich, Institute of Pathology, München, Germany; ¹⁰German Cancer Consortium (DKTK), Partner Site Munich, München, Germany; ¹¹Helmholtz Zentrum München, German Research Center for Environmental Health GmbH, Research Unit Radiation Cytogenetics, Neuherberg, Germany; ¹²Faculty of Physics, Ludwig-Maximilians-Universität München, Department of Medical Physics, München, Germany

Show Affiliations

Purpose or Objective

HPV-status is a known prognostic factor for therapy outcome of head and neck squamous cell carcinomas.

Material and Methods

In this study we used transcriptome data of 348 patients with known HPV-status to build and compare the performance of regression-based and 2D Convolutional Neural Network (CNN) models for HPV prediction. For the CNN, transcriptome data was reorganized as 2D treemap images representing MSigDB Hallmark pathways. The treemaps were built three times in three different ways to assess the stability of the 2D-CNN. The features for the linear regression model were selected using Lasso on the training set.

Results

Applied on an unseen testing set comprising 25 % of the full patient dataset, the CNN achieved test ROC-AUCs/PR-AUCs of 0.95/0.87, 0.93/0.82 and 0.93/0.81 for the three variants of input treemaps respectively, while the regression model achieved a ROC-AUC/PR-AUC of 0.92/0.81.

Conclusion

Therefore, we conclude that accurate predictions of HPV-status can be made with both models. However, the advantage over linear regression is that the deep learning model allows for functional interpretation through visualization of saliency maps computed using the Grad-CAM method.