In many different clinical contexts, elements of interest can be extremely small and their detection and individual characterisation essential to be used as potential imaging biomarkers. This is particular the case in the oncology domain, where the detection of malignant masses at an early stage can be critical to patient care strategy and ultimate outcome. Such problem also occurs in the field of neurology and notably when investigating very small age related neurovascular changes such as enlarged perivascular spaces or lacunes. Their manual annotation is extremely challenging, time-consuming and prone to inter and intra-rater variabiity which makes the development of automated solutions very appealing.
When classical segmentation solutions are inadequate to the detection of such objects, refined, targeted engineering solutions can be designed. In the field of object detection, one of the key aspect is to adequately limit the areas in which to look for the component of interest to reduce the dramatic imbalance between foreground and background.
While it may be enough to find an area of interest in which each element of interest is contained, one may want to further segment each of these objects, consider different classes of objects in the same image or characterise elements with very different shapes and volumes.
In this context, deep learning based frameworks are investigated to solve such issues and help to the creation of new ways of providing robust, accurate and reproducible measures for further clinical use.
Morrell, S., Wojna, Z., Khoo, C.S., Ourselin, S. and Iglesias, J.E., 2018. In Image Analysis for Moving Organ, Breast, and Thoracic Images (pp. 64-72). Springer, Cham.