The paper presented an automatic and precise single-stage classification method by comparing three state-of-the- art networks to select the model fulfilling both the best performance and the fastest computation for clinical use, which successfully predicts whether the patient has AAA or not with corresponding presence probability and the CAM technique to prove feature extraction of the chosen model accurately focuses on the AAA region in the training process.
With the help of transfer learning with ImageNet pre-trained weight, our best performing model DenseNet-121 on the clinical AAA dataset achieved an accuracy of 99.6%, a precision of 97%, a recall of 99.8%, a F1-score of 98.2%, and a ROC-AUC of 99.7% with comparative low complexity.
In conclusion, the model with precise detection of the AAA and related existence probability is presented, and we hope our trained model can be applicable to assisting the diagnosis of radiologists and the daily routine examination of patients for the purpose of detecting the presence of AAA timely before the rupture and saving more people from this life-threatening cardiovascular disease.