Chest Radiograph Interpretation with Deep Learning

Chest radiographs are one of the most commonly used diagnostic modalities in healthcare due to their
\neffectiveness in detecting conditions related to the thoracic region. However, the increasing demand
\nfor radiological services, coupled with a shortage of radiologists and the potential for diagnostic errors,
\ndemands innovative solutions. The integration of artificial intelligence (AI) and deep learning techniques
\noffers a promising approach to support radiology professionals and enhance the diagnostic accuracy and
\nefficiency of the diagnostic process. This thesis addresses the challenge of improving the performance
\nof deep learning models for multi-label classification of chest radiographs using the CheXpert dataset. It
\nfocuses on both model-centric and data-centric methods, specifically techniques such as Gaussian Mixture
\nModels (GMM) for relabeling uncertain labels, multi-scale template matching for focused learning, a
\ncustom pooling layer for better feature extraction, and Sequential Multi-Label Enrichment (SMLE) for
\nimproved detection of coexisting conditions. The primary objective is to enhance the reliability and
\nperformance of AI models in chest radiograph interpretation by developing new model- and data-centric
\nmethods. This research aims to improve uncertain label handling, overall classification performance,
\nand the detection of multiple coexisting conditions.
\nThe study employs convolutional neural networks (CNNs), DenseNet121, and Swin Transformers to
\ninvestigate the effects of excluding versus relabeling uncertain labels. It also evaluates the efficacy of a
\ncustom pooling layer for classification performance and the effectiveness of theSMLEDenseNet model on
\nco-existing conditions. Significant performance improvements are demonstrated when uncertain labels
\nare appropriately handled using GMM. The custom pooling layer significantly enhances the classification
\nperformance of the model, and the SMLE DenseNet model outperforms the baseline DenseNet121 in
\ndetecting co-existing conditions. Finally, the study also examines radiologist’s shift in perception towards
\nthe integration of AI in clinical practice through pre- and post-presentation questionnaires. The findings
\nindicate a generally positive attitude shift towards AI. Despite concerns regarding the deskilling of new
\nradiology professionals, radiologists recognize the potential of AI models to enhance the efficiency of
\nradiology departments and patient care.
\nThis research contributes insights into the practical implementation of AI in medical imaging. It
\nshows that advanced techniques can significantly improve the diagnostic performance and efficiency of
\nAI models. The findings also emphasize the importance of addressing radiologist’s concerns to ensure
\nthe successful integration of AI into clinical practice.