usama13o/Penumothroax-Semgntatoin---Undergrad-Dissertation

Neural Networks have revolutionised the healthcare industry by introducing Deep Learning models that reach human-level accuracy. In this work, a pipeline implementation is explored that tries to enhance the accuracy of Deep Learning models in the detection and segmentation of pneumothorax from a set of chest radiographs. This is done in the hopes of facilitating future applications in the medical field. By cascading a CNN and a U-Net, we ensure that the CNN will filter out all the cases with no pathology, i.e.’Normal’ cases, leaving the U-Net (trained on positive cases only) to focus on arranging a set of filters for the separation of pathology from the infected lungs. The models are cross validated on the same data to ensure consistency. The proposed method demonstrates higher accuracy in both segmenting and detecting the pathology. By implementing it with more complex architectures and integrating in the domain knowledge of radiologists, this methods can be applied in conjunction with other applications to rapidly triage and prioritise cases for the presence of anomalies.

Jupyter NotebookAGPL-3.0