Strategising Template-Guided Needle Placement for MR-targeted Prostate Biopsy

Clinically significant prostate cancer has a better chance to be sampled during ultrasound-guided biopsy procedures, if suspected lesions found in pre-operative magnetic resonance (MR) images are used as targets. However, the diagnostic accuracy of procedure is limited by operator-dependent skills and experience sampling targets, sequential decision making process that involves navigating an ultrasound probe placing series needles for potentially multiple This work aims learn reinforcement learning (RL) policy optimises actions continuous positioning 2D views with respect guiding template, such MR targets can efficiently sufficiently. We first formulate task Markov (MDP) construct environment allows targeting performed virtually individual patients, based on their anatomy derived from images. A patient-specific thus optimised, before each procedure, rewarding positive MDP environment. Experiment results fifty four patients show proposed RL-learned policies obtained mean hit rate (HR) 93% average core length (CCL) 11 mm, which compared favourably two alternative baseline strategies designed humans, without hand-engineered rewards directly maximise these clinically relevant metrics. Perhaps more interestingly, it RL agents learned were adaptive lesion size, where spread was prioritised smaller lesions. Such strategy not been previously reported or commonly adopted clinical practice, but led overall superior performance, achieving higher HR (93% vs 76%) measured CCL (11.0 mm 9.8 mm) when intuitively strategies.