Abnormality detection and characterisation in neuroimaging using deep learning

该项目的目标是开发一种决策工具，使用深度学习识别MRI脑部扫描的异常，用于分类和识别先前报告的扫描中的假阴性。该项目将使用来自三个不同地点的> 100，000个最低处理的MRI扫描的回顾性数据集，这些数据集由顾问神经放射学家和经过验证的NLP算法组合标记。我们使用有监督的CNN（DenseNet-121）在这些数据上建立并验证了一个分类模型，具有很高的准确性（AUC 0.943）。一个目的是通过重新审视一些自动NLP标记来扩展这项工作。目前还不清楚如何训练和利用基于BERT的模型来自动标记放射学报告，特别是英国的神经放射学报告。大多数现有的方法依赖于在大型预训练语言模型上微调分类头。初步工作似乎表明，从头开始训练的模型或那些继续在神经放射学报告上进行预训练的模型将优于这些模型。在此之后，有必要证明标签分类的改进是否会导致下游图像分类的有意义的差异。该项目的另一个目的是通过使用深度学习的规范建模来消除异常。规范建模试图通过对健康对照进行训练来获得健康解剖结构的表示，任何显著偏离都将被视为异常。这种方法与监督CNN方法有根本的不同，后者试图生成最能区分健康和异常检查的特征。

The objective of this project is to develop a decision-making tool that identifies abnormalities on MRI brain scans using deep learning, for triage and identifying false negatives in previously reported scans. The project will use a retrospective dataset of >100,000, minimally processed, MRI scans from three different sites, which have been labelled by a combination of consultant neuroradiologists and a validated NLP algorithm. A classification model has been built and validated on this data, using a supervised CNN (DenseNet-121), with high accuracy (AUC 0.943).One aim is to extend this work by revisiting some of the automatic NLP labelling. It is currently unclear the optimal way to train and utilise BERT-based models for the automatic labelling of radiology reports, specificially to neuroradiology reports in the UK. Most existing approaches rely on fine-tuning a classification head on large pre-trained language models. Preliminary work seems to suggest that models trained from scratch or those that continue pretraining on neuroradiology reports would outperform these models. Following this, it would be necessary to prove whether improvements in label classification led to a meaningful difference in downstream image classification.Another aim of this project is to characterise abnormalities by normative modelling using deep learning. Normative modelling seeks to gain a representation of healthy anatomy by training on healthy controls, any significant deviation from which would be considered abnormal. The approach is fundamentally different from supervised CNN approaches, which instead attempts to generate features that best separate healthy and abnormal examinations.