Needle-based spine drug delivery is a procedure that is commonly used to reduce patients’ pain. This is achieved by injecting either local anaesthetic or steroid at the site of pain in order to directly reduce the pain or to decrease inflammation and swelling. The preferred target is defined from the patients’ symptoms. Common targets include the surrounding space of the spinal cord (epidural), the bones’ joins (facet), the nerve root or the intervertebral disks (discography). All these procedures require precise identification of the source of pain and careful planning of the needle trajectory. Both the identification of the target site and the planning are currently performed by experienced surgeons or radiologists, without any planning tool apart from a volumetric CT at the beginning of the procedure and multiple contrast enhanced CTs during the injection.
One of the key planning challenges during injections is that the patients’ pain usually occurs from abnormal stress between joints or broken bones. Such patients’ often have abnormal anatomy or implants which may require the surgeon to deviate from standard trajectory approaches.
For this project, we propose to develop a planning tool to support clinicians undertaking these procedures. This will be achieved by focusing on three main tasks.
First, the student will create a tool to automatically localise the target area. Using natural language processing, clinical reports that contains the symptoms will be processed to determine the vertebrae level as well as the side(s) that needs treatment. From this information, reinforcement learning strategies will be used to label the target on CT in the presence pathology. Reinforcement learning has been shown to be an effective tool to identify spine anatomy such as vertebrae, disk and canal. This task is otherwise challenging due to the similarities from one area to another. In 70% of the available retrospective cases, MRI of the neck area are also available. This will be used to augment the CT images in order to better identify the nerves and other non-bony tissues.
Second, using statistical tools, the retrospective data will be analysed to define, in 3D rather than 2D, the needle trajectory properties that differentiate good patient outcomes from bad patient outcomes. Example properties could potentially include minimal distance to the bone, distance to structure at risk, needle length or needle bending amongst others. These markers will then be used to implement an automated planning tool. The planning tool will leverage on the software infrastructure of ongoing projects for keyhole tool placement, especially EpiNav.
Last, the student will assess the feasibility to develop a learning-based prognosis tool inferring from needle trajectory. This would potentially enable to later optimise the needle placement so that it maximises the benefit to the patient.