Imaging to predict checkpoint inhibitor outcomes in advanced melanoma

This thesis investigates the predictive value of computed tomography (CT) and histopathology based predictors for outcomes of checkpoint inhibitor treatment in patients with advanced melanoma. The overarching goal is to identify accurate predictors based on routine diagnostic modalities that can guide clinical decisions, thereby mitigating unnecessary toxicity and healthcare costs.

Chapter 1 introduces the role of checkpoint inhibitor treatments in managing advanced melanoma, emphasizing the significant improvements in patient survival rates. It also highlights the challenge of predicting treatment outcomes due to the variable response among patients, serious toxicity and high costs associated with these treatments.

Chapter 2 presents a systematic review of previous literature on imaging biomarkers for response and survival in checkpoint inhibitor treatments across cancer types. It found evidence supporting several imaging biomarkers, such as tumor burden and liver metastases, but notes the limited added value of baseline FDG-PET parameters. Radiomics and radioactive drug labeling are highlighted as promising yet methodologically challenging approaches.

We assessed the predictive value of CT radiomics in chapter 3. These radiomics are visual characteristics such size, roundness and texture, which together capture the appearance of a metastatic lesion on CT imaging. A machine learning model was trained to predict clinical benefit based on the extracted features. While this model achieved moderate predictive accuracy (AUROC = 0.61), it did not significantly outperform a clinical model (AUROC = 0.65). An overlap in information between radiomics and clinical predictors can explain why radiomics does not provide additional value for predicting checkpoint inhibitor outcomes in melanoma.

Chapter 4 investigates deep learning models trained on CT images of metastatic lesions for predicting checkpoint inhibitor treatment outcomes. The hypothesis explored in this chapter was that this method will improve over the method in chapter 3, because it is not limited by the choice of extracted features. Results were similar to those of chapter 3: the model achieved a significant predictive value (AUC = 0.61) but did not improve over a model of known clinical predictors (AUC = 0.64). An overlap in learned information appears to explain this lack of improvement.

Chapter 5 explores the association between body composition metrics derived from CT scans and treatment outcomes. For 1471 patients, data was collected on pretreatment body mass index, in addition to the CT-derived metrics skeletal muscle index and density, and subcutaneous and visceral adipose tissue. Results suggest that underweight BMI, lower skeletal muscle density and higher visceral adipose tissue index associated with worse outcomes, which is independent of known clinical predictors.

The preliminary study presented in chapter 6 investigates deep learning to analyze histopathology images for predicting treatment outcomes. The investigated deep learning model works by splitting the whole-slide histopathology image into square, non-overlapping patches, extracting features from these patches using pretrained deep learning models and training a classifier to predict treatment outcomes based on these features. This model did not perform significantly better than random for predicting treatment outcomes in both primary (AUC = 0.50) and metastatic samples (AUC = 0.54). Although further research is required, this preliminary result suggests that predicting checkpoint inhibitor treatment outcomes based on histopathology imaging is a challenging task.

The discussion emphasizes the challenging nature of predicting immune response to cancer and suggests that advancements in understanding cancer immunology and high-resolution measurements are crucial for future improvements in prediction accuracy. Recommendations for future research include leveraging new methodologies like spatial transcriptomics and ensuring data-driven techniques are evaluated against known predictors. Despite identifying several predictors that add to existing knowledge, the thesis acknowledges that current predictive accuracy is insufficient to withhold checkpoint inhibitor treatment, although it can be used to inform shared decision making.