Solid tumour segmentation via principal axis estimation using weakly supervised adversarial deep learning

Deep learning has dominated the medical imaging literature in recent years. Convolutional neural networks (CNNs) have been particularly successful at learning to perform highly complex tasks in a matter of hours that require many years of training on the part of a human annotator. A critical omission from this narrative is the volume of manual annotation required in the first place in order to train such models. Despite ground- breaking innovations in machine vision over the past few years, much of the attention has been focused on diseases or modalities for which large annotated datasets are readily available. While image-level annotation may be achieved in a reasonable timeframe, per-pixel annotations are considerably harder to obtain at scale.

Segmentation is a perennial component of many image-based diagnosis pipelines, where the boundaries of anatomical structures or anomalies are delineated to enable calculation of morphological features (e.g. size & shape), monitoring of growth/shrinkage, and planning for surgical or therapeutic procedures. Invariably, there is little room for error in this process. Manual delineation of medical images remains the gold standard in many disciplines, requiring highly laborious and painstaking efforts on the part of expert annotators. In light of increasing demands for AI-based solutions in healthcare, there has been a shift towards techniques that can leverage noisy labels.

Weakly supervised learning is a paradigm where ground truth data are provided in the form of less-than-perfect labels. This imperfection can often be due to inherent noise in the annotation process, but is increasingly by design as a means of reducing the annotation burden. For medical image segmentation, weak labels may take the form of data that are already collected as part of routine clinical care but fall short of a complete segmentation. For example, the measurement of a tumour's perpendicular diameters (i.e. principal axes) is often performed to estimate cross-sectional area over time to monitor treatment response. Techniques such as RECIST (Eisenhower et al. 2009) and RANO (Wen et al. 2010) are considerably less time consuming to perform than a complete segmentation, but still require comprehensive knowledge of tumour presentation and morphology. This motivates the development of automated methods that can learn how to perform such measurements and transfer their knowledge to perform segmentation without paired ground truth data.

This PhD project aims to develop a medical image segmentation approach based on weakly supervised and adversarial deep learning. CNNs will be trained to perform bidimensional measurements from medical images: RECIST (for lung tumours) and RANO (for brain tumours). A backbone network based on DenseNet (Huang et al. 2017) will be used for low-level feature learning, which is connected to specialised layers for principal axis and centroid estimation. The specialised layers will then be fused into a final segmentation prediction layer. Adversarial learning (Goodfellow et al. 2014) will be used to promote the generation of labelmaps that appear visually similar to those from an external dataset. In principle, these data may have originated from a different patient population or imaging modality. The method will be developed and validated using two publicly available datasets; BraTS (multi-sequence MR images of glioma) and TCIA (CT images of lung tumours). Bidimensional measurements will be synthetically generated from labelmap data using previous methods (Chang et al. 2019), with variations introduced to mimic the inter-rater variability observed among clinicians. An in-depth evaluation will be conducted to determine the quantity and quality of weakly supervised data needed to achieve competitive segmentation performance. Clinical support and data for validation will be sought from Lincoln County hospital.