Abstract
Background
Lung cancer is the most common cancer in Hong Kong, with 85% of lung cancers are non-small cell lung cancer (NSCLC). Radiation therapy is the mainstream treatment strategy with intensity-modulated radiotherapy (IMRT) technique and stereotactic body radiotherapy (SBRT) technique. With the advancement of artificial intelligence (AI) technologies, different new machine learning systems have emerged with higher computing power for more complicated data analysis.
Method
488 NSCLC patients who received radiotherapy in Prince of Wales hospital were included in this study. 107 radiomics were retrieved with planned target volume (PTV) from CT images using Slicer (V.4.11.20210226) with the PyRadiomics extension. Random Forest from R was used to build the prognosis prediction model. Patients from 2014-17 was used as training and validation data with 70:30 ratio. Patients from 2018-19 was used as testing data. The performance of random forest model and whether radiotherapy technique with better prognosis can be recommended was assessed using Receiver Operating Characteristic Curve (ROC), with sensitivity (Se), specificity (Sp), accuracy (A) and area under the curve (AUC).
Result
The performance of the random forest model was good, with Se 85%, Sp 38%, A 61.6% and AUC 72%. The ability to recommend a radiotherapy technique with better prognosis was similar, with Se 85%, Sp 42%, A 62.75% and AUC 72%. The size zone non-uniformity, dependence non-uniformity and busyness played important roles in prognostic predication and technique recommendation.
Conclusion
The random forest model based on radiomics was good for radiotherapy technique recommendation. Further investigation can be performed using radiomics from multiple image modalities, genomic data and more treatment options can be recommended.
Lung cancer is the most common cancer in Hong Kong, with 85% of lung cancers are non-small cell lung cancer (NSCLC). Radiation therapy is the mainstream treatment strategy with intensity-modulated radiotherapy (IMRT) technique and stereotactic body radiotherapy (SBRT) technique. With the advancement of artificial intelligence (AI) technologies, different new machine learning systems have emerged with higher computing power for more complicated data analysis.
Method
488 NSCLC patients who received radiotherapy in Prince of Wales hospital were included in this study. 107 radiomics were retrieved with planned target volume (PTV) from CT images using Slicer (V.4.11.20210226) with the PyRadiomics extension. Random Forest from R was used to build the prognosis prediction model. Patients from 2014-17 was used as training and validation data with 70:30 ratio. Patients from 2018-19 was used as testing data. The performance of random forest model and whether radiotherapy technique with better prognosis can be recommended was assessed using Receiver Operating Characteristic Curve (ROC), with sensitivity (Se), specificity (Sp), accuracy (A) and area under the curve (AUC).
Result
The performance of the random forest model was good, with Se 85%, Sp 38%, A 61.6% and AUC 72%. The ability to recommend a radiotherapy technique with better prognosis was similar, with Se 85%, Sp 42%, A 62.75% and AUC 72%. The size zone non-uniformity, dependence non-uniformity and busyness played important roles in prognostic predication and technique recommendation.
Conclusion
The random forest model based on radiomics was good for radiotherapy technique recommendation. Further investigation can be performed using radiomics from multiple image modalities, genomic data and more treatment options can be recommended.
| Original language | English |
|---|---|
| Pages (from-to) | 1397-1397 |
| Number of pages | 1 |
| Journal | Annals of Oncology |
| Volume | 34 |
| Issue number | S3 |
| Early online date | 14 Nov 2023 |
| DOIs | |
| Publication status | Published - Nov 2023 |
