B-044 A Statistical and Rule-Based Approach to Automating Glioblastoma Detection

Abstract Background Glioblastoma Multiforme (GBM) is one of the most aggressive brain tumors to diagnose and treat. Due to the higher prevalence of glioblastoma in recent years, it is critical that it’s detection can be automated in order to facilitate patient care. Furthermore, automation would decrease time for detection of glioblastoma from frozen sections during surgical resection. Two major problems hinder the automated detection of this pathology which include the explainability of the algorithm when arriving at diagnostic conclusions and the immense library of images required to train these systems to be highly accurate, both issues pertaining to machine-learning (ML) based approaches. Methods Image files were extracted from their respective folders (GBM vs control). For feature extraction, the extract_nuclear_features function was used to process each image. Extracted features included: Area, Perimeter, Eccentricity, Mean Intensity, Solidity, Major Axis, and Minor Axis. Baseline Calculations included the average nuclear features for both groups. For each GBM image, the Log2 fold change (log2FC) was compared to normal tissue. P-values were generated using a t-test between the sample and normal tissue to highlight significant change detection. A Volcano Plot was generated to display the significant features of GBM nuclear morphology visualizing log2FC vs. -log10(p-value) for all features to identify significant changes. For the model prediction, two criterias needed to be met for GBM detection. The first was a similarity score with the percentage of features matching GBM patterns (threshold: 70%). The second was the Key Diagnostic Features Match which required a 100% match for key features (Eccentricity, Solidity, Area, Mean Intensity). Results The test results for glioblastoma positive images with random magnifications (n=5), our model detected 2 (40%) as being glioblastoma positive while the other 3 (60%) were inconclusive due to too high magnification and/or too few nuclear presence for pattern detection. A new set of glioblastoma positive images with standardized magnification (n=5) showed 80% accuracy at detecting glioblastoma positive images with only one image that had atypical pattern on histology resulting in no detection. Conclusion Our results demonstrate that a statistical rule-based approach is able to detect GBM with high accuracy given that the model is properly trained with the parameters being tested. Our model reached 80% accuracy with less than 100 images while the industry standard looms somewhere around 85%-95% accuracy after being trained with as many has 50,000 images. Therefore, a simpler model with more directed training is able to compete with much more complex ML-based approaches.