A Simple and Robust Machine Learning Method for Segmentation of Thoracic Computed Tomography

Machine learning being used as a tool in the field of healthcare offers us a great advantage, with already implemented automated processes like segmentation when it comes to cancer patients and treatments. However, these are not widely used as they are computationally costly and do not give a clear concise validation of whether its predictions are correct or not, along with no indications of the limitations or robustness of the model itself. In this work, we aim to make a faster, less complicated model to segment disease/infection within the lungs, along with giving us a score on its uncertainty in its predictions with a look into its limitations with the intention to apply this model to cancer patients with early stage lung cancer in the future. This will allow clinician’s to make a quick decision on how much to rely on this as a tool for that singular prediction in that moment. The lower the score, the worse the prediction in which the clinician will take over, or if the prediction obtains a high score, clinicians can have more trust and take its results into consideration as it gives a level of confidence behind its prediction. This tool could be the key to shortening wait times in the NHS and freeing up time for clinicians to increase survival rates.