Improving the Algorithms required for L3 level Lung Cancer Detection
In order to ensure reproducibility, the initial code required extensive restructuring prior to processing the data. This ensures that future users will be able to utilize the algorithm with minimal updates to the published code
CODE ORGANIZATION & IMPROVEMENTS
For ease of processing, the analysis code was broken into two sections: Data-Pre-processing and L3 Identification. Additional code ensured that the outputs were placed in clearly defined locations rather than simply being output to the working directory.
Additional code ensured that the outputs were placed in clearly defined locations, rather than simply being output to the working directory. We also create an index of functions and classes used in the Sarcopenia-ai model, to aid in future troubleshooting.
In addition to workflow improvements, the team also addressed detection improvements. To change the algorithm’s precision and recall, image pre-processing was modified. We avoided tuning the FCNN parameters because it was already trained.
Pre-processing parameters are easily accessible and changeable in the code. The team changed the Hounsfield unit (HU) parameter, which refers to the absorption and attenuation of x-rays on tissue.
We raised the minimum HU from 100 to 200, which can be interpreted as changing the grayscale values of the image.
We will be sharing evidence of improvement, lessons learned, and future work through the next weeks. Please follow us on LinkedIn to stay tuned.