Purpose: Despite the recent growth in artificial intelligence (AI), the “black-box” nature brings a need for “explainable” AI (XAI) to better understand system performance and limitations. We apply this concept to an AI-based system for lung nodule detection in CT, which can be decomposed into steps of (1) candidate generation (CG), and (2) false positive reduction. The specific purpose of this work is to demonstrate the utility of a Genetic Algorithm (GA} in improving performance of CG in an explainable, understandable fashion.
Methods: 1010 CT lung image cases from the Lung Image Data Consortium were used. Nodules >=4 mm were identified based on markings from at least 3 radiologists. A previously validated SN CAD system was modified to enable tunable parameters for nodule CG, encoded as chromosomes in the GA. A training subset was used for progressive training, and a validation subset to select the optimal chromosome. Performance on the test set was measured for SNs: (a) hand-tuned, no GA, (b) GA, basic operations (GA-basic), or (c) GA, complex operations (GA-complex). GA-complex incorporated additional operations, to try to promote GA concepts of diversity (maintaining sufficient genetic heterogeneity in each generation) and elitism (propagation of the higher performing chromosomes). Performance was measured via sensitivity and false positive (FP) rate. Parameter importance was measured through random forest variable importance.
Results: Trained from scratch, GA-complex-optimized CG matched hand-tuned model performance with 90.8% sensitivity with false positive rate greatly reduced to 498 FP/case. Tuning parameters "minimum x-y circularity", "HU-range", and "maximum x-y area" demonstrated the highest impact.
Conclusion: GA-optimization permitted construction of a CG model from scratch that met hand-tuned performance. Further training could enable better performance than human-optimized model parameters. The GA allows examination and deconstruction of chromosomes in the optimization process, to easily follow and manage the evolution of parameters over time.
Funding Support, Disclosures, and Conflict of Interest: UCLA Radiology has a Master Research Agreement with Siemens Healthineers. M. McNitt-Gray receives grant support from Siemens Healthineers and is a Member of the Board of Scientific Advisors for Hura LLC.