Convolutional neural networks (CNNs) excel in medical image analysis, handling tasks like disease classification, tumor segmentation, and lesion detection by extracting local features. In recent developments, transformers have been employed for a wide range of clinical applications in medical imaging, including reconstruction, registration, segmentation, detection, and diagnosis. Multi-modal medical images such as CT, MRI, SPECT, PET with their explicit long-range dependencies, can enhance deep learning (DL) model performance compared to natural images.
Pulmonary diseases like tuberculosis (TB), bacterial pneumonia, and viral pneumonia carry the dual threats of contagion and fatality, posing significant risks to public health. Timely and accurate detection of these illnesses is vital, enabling healthcare professionals to administer prompt and effective treatment. Thus, the effective and precise diagnosis of these diseases holds paramount importance. Conventional diagnostic procedures necessitate extensive manual interpretation, introducing the risk of errors and delays in patient care. In recent years, DL techniques have surfaced as potent tools for automating the classification of pulmonary diseases.
The challenges associated such as limited and data imbalance, data quality, inadequate pre-processing, covariant shift, biases and fairness, labelling error, overfitting, data privacy and regulations, etc. with the medical image classification and segmentation can have a direct relationship on the final performance of any DL model. Due to its composite formation, dealing with CXR images becomes difficult when it is infected, for example, widespread ground-glass opacities and diffuse reticular-nodular opacities. This makes the computerised recognition of pulmonary disease employing CXR imaging a challenging job. Furthermore, the available DL-base model for pulmonary disease classification are lacking to address the local and global relationship of features and challenges of poor data quality resulting in lower accuracy.
By addressing the aforementioned challenges and harnessing the capabilities of artificial intelligence, this PhD. project strives not only to develop an efficient DL-based approach for pulmonary disease classification but also to explore contributions to the broader implementation of cascaded DL models. Through these innovative research endeavours i.e., DL- base cascaded approaches; our goal is to bridge the existing research gap in pulmonary disease classification, particularly in the context of multi-classification challenges.
Methodology:
The research methodology encompasses the following key components:
- Data Collection: Curate comprehensive CXR datasets representing a range of pulmonary diseases, including viral pneumonia, bacterial pneumonia, TB, and normal cases.
- Extensive Pre-processing: Address data quality issues, manage outliers, and employ dimensionality reduction techniques to optimise the dataset for modelling.
- Model Development: Create and implement an AI-driven cascaded DL model tailored specifically for pulmonary disease classification.
- Model Evaluation: Conduct a series of meticulous experiments and performance assessments, followed by iterative model refinement to attain optimal outcomes.
- Performance Metrics: Choose relevant performance metrics, including precision, recall, F1-score, and others, to ensure a comprehensive evaluation of the model's effectiveness.
- Ethical Considerations: Integrate a thorough assessment of ethical and privacy concerns into the research methodology to safeguard sensitive information and adhere to ethical standards.
Applicants should have an honours degree with a 2.1 or above (or equivalent) in Computer Science, Physics or related disciplines. In addition, they should have excellent programming skills in Python in an academic project, good mathematical background and an interest in ML, DL, and intended collaborative AI.
Supervisor: Dr Tariq Rahim
View Tariq Rahim's profile and contact details >