Initial Preprocessing Recognition Trial: Edge-Based Detection
Edge detection primarily consisted of utilizing the open source Pinetools platform for Sobel-Feldman image preprocessing. The overarching protocol for this second phase of experimentation is described in the “Methods and Protocols” section of the document. However, the same data analysis operation (the sampling of 25 images across two broad negative and positive classifiers) was followed through the three preprocessing stages of the project.
Blinded Trials: Bias Elimination and Statistical Validation
The necessity of eliminating bias is a prominent component of all research, since even when preventative measures are taken, researchers have a natural tendency of unintentionally selecting data that enhances their viewpoint. As mentioned previously in the Data Analysis section, all images used were screened for accuracy and high contrast features, but an undetected bias may have been omnipresent in the selection of data. In order to further authenticate the data collected while simultaneously reducing bias, blinded trials were conducted with the intention of negating any form of selective bias. It was decided that thirty images should be selected in a process as hands-off as possible. Bringing in foreign parties to assist with data collection was deemed necessary. Three individuals, each with a different scientific background, were tasked with collecting ten amyloid-beta or tau PET scans which would be tested for senile dementia. Their selection process was guided and moderated by the research team, since a certain degree of instruction was necessary to guide them to the , but the image choice was entirely up to the participants. This
Conclusion
The project amalgamated a series of multi-phased pre-processing and Artificial Intelligent API protocols prior to analysis. Protocols including edge-based detection and gamma enhancement have been key variables in feature detection. Accuracy readings indicate that intense edge-based and gamma detection maximize accuracy scores to nearly 92-100%. Through heat map identification, accuracy falling within 90-100% was achieved through tile and image separation (processed through the IBM Watson Visual Recognition API). The combination of heat mapping and pre-processing techniques revealed and diagnosed positive-negative regions of ambiguous areas of PET trial data.