RESUMO
Supervised machine learning-based medical image computing applications necessitate expert label curation, while unlabelled image data might be relatively abundant. Active learning methods aim to prioritise a subset of available image data for expert annotation, for label-efficient model training. We develop a controller neural network that measures priority of images in a sequence of batches, as in batch-mode active learning, for multi-class segmentation tasks. The controller is optimised by rewarding positive task-specific performance gain, within a Markov decision process (MDP) environment that also optimises the task predictor. In this work, the task predictor is a segmentation network. A meta-reinforcement learning algorithm is proposed with multiple MDPs, such that the pre-trained controller can be adapted to a new MDP that contains data from different institutes and/or requires segmentation of different organs or structures within the abdomen. We present experimental results using multiple CT datasets from more than one thousand patients, with segmentation tasks of nine different abdominal organs, to demonstrate the efficacy of the learnt prioritisation controller function and its cross-institute and cross-organ adaptability. We show that the proposed adaptable prioritisation metric yields converging segmentation accuracy for a new kidney segmentation task, unseen in training, using between approximately 40% to 60% of labels otherwise required with other heuristic or random prioritisation metrics. For clinical datasets of limited size, the proposed adaptable prioritisation offers a performance improvement of 22.6% and 10.2% in Dice score, for tasks of kidney and liver vessel segmentation, respectively, compared to random prioritisation and alternative active sampling strategies.
Assuntos
Algoritmos , Humanos , Tomografia Computadorizada por Raios X , Redes Neurais de Computação , Aprendizado de Máquina , Cadeias de Markov , Aprendizado de Máquina Supervisionado , Radiografia Abdominal/métodosRESUMO
PURPOSE: Magnetic resonance (MR) imaging targeted prostate cancer (PCa) biopsy enables precise sampling of MR-detected lesions, establishing its importance in recommended clinical practice. Planning for the ultrasound-guided procedure involves pre-selecting needle sampling positions. However, performing this procedure is subject to a number of factors, including MR-to-ultrasound registration, intra-procedure patient movement and soft tissue motions. When a fixed pre-procedure planning is carried out without intra-procedure adaptation, these factors will lead to sampling errors which could cause false positives and false negatives. Reinforcement learning (RL) has been proposed for procedure plannings on similar applications such as this one, because intelligent agents can be trained for both pre-procedure and intra-procedure planning. However, it is not clear if RL is beneficial when it comes to addressing these intra-procedure errors. METHODS: In this work, we develop and compare imitation learning (IL), supervised by demonstrations of predefined sampling strategy, and RL approaches, under varying degrees of intra-procedure motion and registration error, to represent sources of targeting errors likely to occur in an intra-operative procedure. RESULTS: Based on results using imaging data from 567 PCa patients, we demonstrate the efficacy and value in adopting RL algorithms to provide intelligent intra-procedure action suggestions, compared to IL-based planning supervised by commonly adopted policies. CONCLUSIONS: The improvement in biopsy sampling performance for intra-procedure planning has not been observed in experiments with only pre-procedure planning. These findings suggest a strong role for RL in future prospective studies which adopt intra-procedure planning. Our open source code implementation is available here .