Deep Learning-Based Medical Image Analysis Methods for Reducing Annotation Costs and Predictive Triage Misdiagnosis

  • Jacob Carse

Student thesis: Doctoral ThesisDoctor of Philosophy

Abstract

Medical image analysis is a critical and challenging field that can be significantly enhanced using deep learning techniques. However, these models require large amounts of annotated data, which can be costly and time-consuming to obtain. Additionally, deep learning models often suffer from overconfidence and poor generalisation, leading to incorrect diagnoses and negative clinical outcomes. The primary objective of this thesis is to address these challenges by proposing methods that reduce annotation costs and improve diagnostic accuracy using deep learning.
      
The first contribution of this thesis is an active learning framework designed to increase annotation throughput for histopathology patches. Histopathology is the gold standard for cancer diagnosis, but it requires manual examination by pathologists, which is labour-intensive and prone to errors. To address this issue, this thesis proposes an active learning framework that selects regions for annotation composed of multiple patches, which is expected to increase annotation throughput. This framework is evaluated with various query strategies for nuclei classification using convolutional neural networks (CNN) trained on small patches containing single nuclei.
      
This thesis proposes a multi-directional modification to the contrastive predictive coding (CPC) method for unsupervised representation learning for histopathology patches. Recent advancements in deep learning have had a significant impact on digital pathology, however a significant challenge is the large amounts of annotated data needed. Unsupervised representation learning aims to learn meaningful and transferable features from unannotated data, which can be useful for downstream tasks such as classification. The proposed method uses an alternative mask to construct a latent context and a multi-directional PixelCNN autoregressor, to learn effective deep feature representations for improved classification accuracy in digital pathology compared to the standard implementation of CPC.
     
The third contribution of this thesis is a study on calibration techniques evaluated on a multi-class dermatology dataset and a binary histopathology dataset. Calibration is critical for medical image analysis, where overconfident or underconfident predictions can have serious consequences for patient care. The study applied the temperature scaling method and alternative calibration metrics to networks trained with one-hot encoding, cross-entropy loss, focal loss, and label smoothing. The findings suggest that temperature scaling of networks trained with focal loss and appropriate hyperparameters demonstrated strong performance in terms of both calibration and accuracy across both datasets.
     
This thesis investigates selective classification methods with asymmetrical misdiagnosis costs for skin lesion images. Selective classification is a decision-making framework that allows a model to reject images when it is uncertain or unconfident, which can reduce the risk of misdiagnosis and improve patient safety. However, most existing selective classification methods assume that all types of misclassification have equal costs, which is not realistic in medical image analysis. This thesis evaluates various methods of uncertainty estimation with neural networks and probability calibration. Additionally, a modification to SelectiveNet, called EC-SelectiveNet, is proposed, which discards the selection head during testing and relies on expected costs to make decisions. The results demonstrate the advantages of training for full coverage, even when operating at lower coverage, and show that EC-SelectiveNet outperforms other selective classification methods, in both symmetric and asymmetric cost settings.
      
The fifth contribution of this thesis is a study on dataset fine-tuning for skin lesion image datasets. Dataset fine-tuning is challenging for medical image analysis due to the heterogeneity and variability of data sources. This study utilises four diagnostic image datasets, including two locally sourced datasets from NHS Tayside and NHS Forth Valley and two publicly available datasets. The study emphasises the importance of assessing the generalisability of deep learning algorithms for macroscopic skin lesion images in real-world settings and highlights the potential benefits of utilising large public macroscopic datasets for pre-training and fine-tuning.
Date of Award2023
Original languageEnglish
SupervisorStephen McKenna (Supervisor) & Francis Carey (Supervisor)

Cite this

'