ESTRO 2022

Session Item

Radiomics, modelling and statistical methods

Session Code: 7011

Session Type: Poster (digital)

Track: Physics

Feasibility of a novel harmonization method for NSCLC multi-centric radiomic studies

Andrea Botti, Italy

Presentation Number: PO-1773

Abstract

Abstract Title:

Feasibility of a novel harmonization method for NSCLC multi-centric radiomic studies

Authors:

Andrea Botti¹, Marco Bertolini¹, Valeria Trojani¹, Noemi Cucurachi², Mauro Iori¹, Marco Galaverni³, Cinzia Iotti⁴, Paolo Borghetti⁵, Salvatore La Mattina⁶, Niccolò Giaj Levra⁷, Matteo Sepulcri⁸, Federico Iori⁹, Patrizia Ciammella¹⁰

¹AUSL - IRCCS Reggio Emilia, Medical Physics, Reggio Emilia, Italy; ²Università di Modena e Reggio Emilia, Physics Department, Modena, Italy; ³Azienda Ospedaliero-Universitaria di Parma, Radiation Therapy Department, Parma, Italy; ⁴AUSL - IRCCS Reggio Emilia, Radiation Therapy Department, Reggio Emilia, Italy; ⁵Spedali Civili di Brescia, Radiation Therapy Department, Brescia, Italy; ⁶Spedali Civili di Brescia, Radiation Therapy, Reggio Emilia, Italy; ⁷IRCCS Ospedale Sacro Cuore Don Calabria, Radiation Therapy, Verona, Italy; ⁸Istituto Oncologico Veneto, Radiation Therapy, Padova, Italy; ⁹Università degli studi di Modena e Reggio Emilia, Radiation Therapy, Modena, Italy; ¹⁰AUSL - IRCCS Reggio Emilia, Radiation Therapy, Reggio Emilia, Italy

Show Affiliations

Purpose or Objective

This work aims to develop a predictive radiomic model using an innovative harmonization technique to evaluate patients affected by non-small cell lung cancer (NSCLC), using simulation CT and PET/CT.

Material and Methods

106 patients were enrolled from six centers within a research project. Radiomic biomarkers, calculated with pyRadiomics software, were evaluated in the volumes selected by the radiotherapy physicians using a validated protocol. Segmentations were placed in the contralateral healthy lung and shifted by 3 and 6 mm in 6 directions to assess the variability and robustness of the radiomic features. The resulting statistical data were used to create harmonized models according to an in-house method developed to reduce the bias caused by the different acquisition protocols used by the participating institutions. In addition, machine learning techniques capable of predicting the probability of overall disease progression at two years were evaluated. 68 patients from two centers were used in the training phase of the model, using a 10-fold cross-validation strategy. 38 patients from the other four centers were used for the external validation of the model.

Results

This harmonization method was able to make the feature distributions in the different centers comparable with each other. Out of the 4506 features from the three modalities, five were chosen using the LASSO technique as feature selection to construct the radiomic predictive models. The three models with the highest accuracy were linear SVM, quadratic SVM, and bagged trees. For the train dataset, the following AUC confidence intervals were obtained for linear SVM, quadratic SVM and bagged trees, respectively: [0.73−0.92], [0.75−0.96] and [0.73−0.90] for harmonized features; while for non-harmonized features they were [0.65 − 0.89], [0.83 − 0.96] and [0.73 − 0.95]. For the external validation dataset, the AUC results were: [0.59 − 0.77], [0.70 − 0.85] and [0.40−0.75] for harmonized features; while for non-harmonized features they were [0.33−0.71], [0.39−0.70] and [0.38−0.76].

Conclusion

In conclusion, our proposed harmonization applied to the quadratic SVM model allows for higher AUC for both the training dataset and the external validation dataset. This result confirms the strength of our harmonization method, highlighted by the smaller difference between training and validation AUC values.