The path towards improving oncological outcomes and quality of life while reducing costs in cancer care is directly linked to personalized medicine. Since some imaging-relevant features are not discernible to the human eye, machine learning (ML) systems are the last resort to identify prognostic and predictive features, allowing clinicians to select the best treatment for a particular patient and oncological characteristics. Two types of ML systems could be used to identify quantitative predictive features and distinguish treatment responders from non-responders: (hand-crafted) radiomics and deep learning. Radiomics approaches, which rely on quantitative measures of morphological, phenotypic and textural characteristics, have demonstrated to be successful for treatment outcome prediction in several clinical scenarios (e.g., immunotherapy response in melanoma and non-small-cell lung cancer). However, radiomics relies on manual or automatic delineations of regions of interest and generates features that are not dependent on the clinical end goal, possibly missing clinically-relevant information. On the other hand, deep learning approaches usually do not require any expert delineation effort and, in principle, can learn any interesting semantic representation of the data adapted to the task, but require a significant number of training samples to generalize reasonably well and lack transparency/interpretability. Therefore, intensive fundamental and applied research has to be conducted to increase the generalizability and interpretability of deep learning models in the context of treatment outcome prediction. Moreover, treatment outcome prediction is only one side of the problem. Also, for treatment outcome assessment, there is a relevant role to be played by ML systems. Currently, response evaluation criteria, like RECIST or RANO are standard metrics to evaluate therapy response. However, the existing metrics do not account for all visually-observable clinical information like immune-related toxicities and cancer-related complications. As such, a gold-standard comprehensive quantitative approach driven by ML is promising and still missing. For many years, the MICCAI community has provided novel diagnosis algorithms to help radiologists improve their decision-making and optimize their workflows. However, little has been done and explored in the context of treatment outcome prediction, currently a main focus of the radiology and radiation oncology communities. Thus, this workshop aims at calling the attention of the MICCAI community, fostering discussions and the presentation of ideas to tackle the many challenges and identifying opportunities related to the topic of treatment response assessment and prediction based on radiology data by means of machine learning techniques.
Therefore, the main purposes of this workshop are:
- To bring the MICCAI and ESR communities together, introducing the challenges/opportunities related to the topic of treatment response assessment and prediction, taking into account both the AI and clinical perspectives on the topic.
- To understand the state of the art of this field. This will be achieved through the submitted manuscripts and the invited keynote speakers.
- To join researchers in this field and to discuss the issues related to it and future work.
- To understand the difficulty and relevance of using machine learning systems in treatment outcome prediction and assessment based on Radiology data.