Abstract
1.1.1 BackgroundRadiomic and deep neural network models yield detailed quantitative descriptions of tumour shape, intensity and texture in medical images, presenting an opportunity to improve clinical predictions of prognosis, therapeutic response and tumour genotype. However, at the time of writing, translation of imaging models into clinical practice remains slow. Concerns have arisen regarding prospective reliability and generalisability in the clinical environment. The image feature space is high-dimensional relative to the typical sample size of clinical cohorts, presenting potential issues for model fitting, feature selection and biological interpretation. Furthermore, image data distributions vary according to external factors including imaging equipment and acquisition parameters. There is a clinical need to improve the reliability of imaging models through interpretable design and appropriate validation.
1.1.2 Objectives
The following objectives were addressed in nine studies:
1. Quantify compliance with best practices in design and reporting of convolutional neural network models for radiological cancer diagnosis according to the Checklist for Artificial Intelligence in Medical Imaging (CLAIM).
2. Develop a convolutional neural network (CNN) image classifier architecture which is interpretable by design using a weakly-supervised segmentation framework for application to lung tumour detection on computed tomography (CT) images.
3. Review the current state of the art in radiomic models for application to oesophageal adenocarcinoma, appraising design and reporting standards and identifying frequently proposed radiomic features.
4. Develop prognostic models for oesophageal adenocarcinoma with clinical predictors and previously proposed pre-treatment CT features.
5. Identify CT radiomic associations with genes of known prognostic significance in oesophageal adenocarcinoma treated with neoadjuvant chemoradiotherapy.
6. Model mean tumour and surface enhancement for prediction of tumour regression grade in oesophageal adenocarcinoma in pre- and post-treatment CT images.
7. Perform Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) type 4 validation study of three previously published CT radiomic radiomic models for prediction of programmed death ligand-1 (PD-L1) expression in non-small cell lung cancer (NSCLC).
8. Evaluate the feature selection performance of penalised regression models by simulation of low-sample size, high-dimensionality conditions using large sample
gene expression data.
1.1.3 Methods
1. Systematic review of studies applying convolutional neural network models to radiological cancer diagnosis from 2016 to 2020 performed by two independent reviewers. Compliance with Checklist for Artificial Intelligence in Medical Imaging (CLAIM) assessed.
2. Weakly-supervised Unet architecture (WSUnet) pipeline developed, learning lung tumour segmentation from image-level data labels, generating voxel probability maps through a Unet and applying global max-pooling to train with image-level loss. Comparison with current model interpretation techniques and clinician preference survey performed.
3. Reviewed research in oesophageal adenocarcinoma radiomics between 2016 and 2022, appraising design and reporting standards according to the TRIPOD score and Radiomics Quality Score (RQS).
4. Modelled 3-year overall survival in oesophageal adenocarcinoma treated with chemoradiotherapy using clinical variables (age, clinical tumour stage (cT-stage), clinical nodal stage (cN-stage)) and previously proposed CT radiomic features in a prospective multicentre Oesophageal Cancer Clinical and Molecular Stratification (OCCAMS) cohort.
5. In OCCAMS cohort, screened CT features for variability with respect to scanner manufacturer and study institution. Using robust features, screened for association with genes of known prognostic significance in oesophageal adenocarcinoma.
6. In OCCAMS cohort, modelled histopathological outcomes following neoadjuvant chemoradiation using clinical variables and quantitative imaging variables extracted from pre- and post-treatment CT.
7. Screened published CT radiomic PD-L1 prediction models for reproducibility using RQS. Standardised radiomic features according to publicly available data and evaluated models’ discrimination and calibration on multi-institutional NSCLC cohort.
8. High-dimension, low-sample size conditions simulated by subsampling publicly available gene expression datasets. Comparison of regression-based feature selection with L0, L1, L2 norm penalties using large-sample test partitions for data-driven validation performed.
1.1.4 Results
1. Many studies did not meet the guidelines for current design and reporting. Compliance correlated positively with publication year and journal H-index. 2. WSUnet learned to localise tumour precisely at voxel-level, learning only from image-level labels, outperforming the best comparator methods. 3. Radiomic feature proposals for oesophageal adenocarcinoma modelling varied between studies, with few features being proposed consistently. Studies rarely validated previously proposed radiomic models or features. 4. A model of clinical features and previously proposed radiomic features yielded the greatest discrimination of 3-yr survival in oesophageal adenocarcinoma. Target test sensitivity was achieved by clinicoradiomic and stage models only. 5. SIRT2 expression correlated negatively with tumour localintensity_Energy in oesophageal adenocarcinoma. This association was not confounded by age, tumour stage or volume. 6. The CT radiomic model predicted oesophageal adenocarcinoma tumour regression grade (TRG) in both pre-treatment and post-treatment timepoints. Clinical features models did not discriminate TRG. A model of cN-stage and post-treatment surface attenuation predicted lymphovascular invasion. 7. Of eleven screened CT radiomic models for PD-L1 prediction, three could be reconstructed from reported information. In standardised radiomic data, one published model matched reported discrimination performance. 8. L1L2-penalised models achieved the greatest cosine similarity with gold-standard coefficients. L0L2 -penalised models explained the greatest proportion of variance in test responses.
1.1.5 Conclusions
1. Imaging studies applying CNNs for cancer diagnosis frequently omitted key clinical information including eligibility criteria and population demographics. Fewer than half of imaging studies assessed model performance on explicitly unobserved test data partitions. 2. Weakly supervised segmentation is a viable approach by which explainable object detection models may be developed for medical imaging. 3. Future research must prioritise prospective validation of previously proposed models to further clinical translation of radiomics in oesophageal adenocarcinoma. 4. Compared to tumour node metastasis (TNM) staging, the multivariate models of prespecified clinical and radiomic variables improved discrimination of 3-year overall survival. 5. A potential prognostic imaging-gene association was identified for oesophageal adenocarcinoma highlighting the potential for localintensity_Energy for aiding risk-stratification. 6. The volume attenuation and surface attenuation of the primary tumour was associated with pathological TRG, and augmented response prediction based on tumour volume and T-stage alone. 7. Reporting of CT radiomic models for PD-L1 tumour proportion score (TPS) prediction was often insufficient for external model reconstruction. Of three validated models, one reproduced reported discrimination performance. A radiomic model of CD274 expression correlates modestly with PD-L1 TPS in external validation. 8. Evaluation of learning algorithms according to observed test performance in external genomic datasets yielded valuable insights into actual test performance, providing a data-driven complement to internal cross-validation in genomic regression tasks.
| Date of Award | 1 Dec 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Vicky Goh (Supervisor), Gary Cook (Supervisor) & Sophia Tsoka (Supervisor) |