Skip to main content
placeholder image

Using machine learning applied to radiomic image features for segmenting tumour structures

Conference Paper


Abstract


  • Lung cancer (LC) was the predicted leading cause of Australian cancer fatalities in 2018 (around 9, 200 deaths). Non-Small Cell Lung Cancer (NSCLC) tumours with larger amounts of heterogeneity have been linked to a worse outcome. Medical imaging is widely used in oncology and non-invasively collects data about the whole tumour. The field of radiomics uses these medical images to extract quantitative image features and promises further understanding of the disease at the time of diagnosis, during treatment and in follow up. It is well known that manual and semi-automatic tumour segmentation methods are subject to inter-observer variability which reduces confidence in the treatment region and extent of disease. This leads to tumour under- and over-estimation which can impact on treatment outcome and treatment-induced morbidity. This research aims to use radiomic features centred at each pixel to segment the location of the lung tumour on Computed Tomography (CT) scans. To achieve this objective, a Decision Tree (DT) model was trained using sampled CT data from eight patients. The data consisted of 25 pixel-based texture features calculated from four Gray Level Matrices (GLMs) describing the region around each pixel. The model was assessed using an unseen patient through both a confusion matrix and interpretation of the segment. The findings showed that the model accurately (AUROC = 83.9%) predicts tumour location within the test data, concluding that pixel based textural features likely contribute to segmenting the lung tumour. The prediction displayed a strong representation of the manually segmented Region of Interest (ROI), which is considered the ground truth for the purpose of this research.

UOW Authors


  •   Vial, Alanna (external author)
  •   Carolan, Martin (external author)
  •   Field, Matthew (external author)
  •   Holloway, Lois (external author)
  •   Miller, Andrew (external author)
  •   Ritz, Christian
  •   Ros, Montserrat
  •   Stirling, David

Publication Date


  • 2019

Citation


  • Clifton, H., Vial, A., Miller, A., Ritz, C., Field, M., Holloway, L., . . . Stirling, D. (2019). Using machine learning applied to radiomic image features for segmenting tumour structures. In 2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2019 (pp. 1981-1988). doi:10.1109/APSIPAASC47483.2019.9023077

Scopus Eid


  • 2-s2.0-85076066587

Start Page


  • 1981

End Page


  • 1988

Volume


Issue


Place Of Publication


Abstract


  • Lung cancer (LC) was the predicted leading cause of Australian cancer fatalities in 2018 (around 9, 200 deaths). Non-Small Cell Lung Cancer (NSCLC) tumours with larger amounts of heterogeneity have been linked to a worse outcome. Medical imaging is widely used in oncology and non-invasively collects data about the whole tumour. The field of radiomics uses these medical images to extract quantitative image features and promises further understanding of the disease at the time of diagnosis, during treatment and in follow up. It is well known that manual and semi-automatic tumour segmentation methods are subject to inter-observer variability which reduces confidence in the treatment region and extent of disease. This leads to tumour under- and over-estimation which can impact on treatment outcome and treatment-induced morbidity. This research aims to use radiomic features centred at each pixel to segment the location of the lung tumour on Computed Tomography (CT) scans. To achieve this objective, a Decision Tree (DT) model was trained using sampled CT data from eight patients. The data consisted of 25 pixel-based texture features calculated from four Gray Level Matrices (GLMs) describing the region around each pixel. The model was assessed using an unseen patient through both a confusion matrix and interpretation of the segment. The findings showed that the model accurately (AUROC = 83.9%) predicts tumour location within the test data, concluding that pixel based textural features likely contribute to segmenting the lung tumour. The prediction displayed a strong representation of the manually segmented Region of Interest (ROI), which is considered the ground truth for the purpose of this research.

UOW Authors


  •   Vial, Alanna (external author)
  •   Carolan, Martin (external author)
  •   Field, Matthew (external author)
  •   Holloway, Lois (external author)
  •   Miller, Andrew (external author)
  •   Ritz, Christian
  •   Ros, Montserrat
  •   Stirling, David

Publication Date


  • 2019

Citation


  • Clifton, H., Vial, A., Miller, A., Ritz, C., Field, M., Holloway, L., . . . Stirling, D. (2019). Using machine learning applied to radiomic image features for segmenting tumour structures. In 2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2019 (pp. 1981-1988). doi:10.1109/APSIPAASC47483.2019.9023077

Scopus Eid


  • 2-s2.0-85076066587

Start Page


  • 1981

End Page


  • 1988

Volume


Issue


Place Of Publication