EfDenseNet: Automated Pulmonary Hypertension Detection Model Based on EfficientNetb0 and DenseNet201 Using CT Images

EfDenseNet: Automated Pulmonary Hypertension Detection Model Based on EfficientNetb0 and DenseNet201 Using CT Images

Pulmonary hypertension (PH) is a chronic and progressive disease. We introduced a novel automated self-organized feature engineering architecture for PH detection, which was trained and refined using a new thoracic CT image dataset. This study’s dataset includes 807 transverse contrast-enhanced CT images from 313 patients, categorized into four groups: Group 1 with 20 mmHg ≤ mean pulmonary artery pressure (mPAP) < 25 mmHg; Group 2 with 25 mmHg ≤ mPAP ≤ 30 mmHg; Group 3 where mPAP > 30 mmHg; and a control group with no PH. Our model consists of four primary stages: (i) generation of features based on combinations from nested patches, (ii) feature selection, (iii) classification and (iv) majority voting. CT images were segmented into nested patches, each being processed through pretrained EfficientNetB0 and DenseNet201 to derive four deep feature vectors, utilizing both the global average pooling and fully connected layers of these networks. These four extracted features underwent combinatorial operations, resulting in 15 feature vectors. Subsequently, these vectors were introduced to neighborhood component analysis, ReliefF, and Chi2 feature selectors. This process yielded 45 refined feature vectors with diminished data dimensions. These selected vectors were then processed through a support vector machine and k-nearest neighbors classifiers, producing 90 predictive vectors. By applying mode-based iterative majority voting to these vectors, an additional 88 voted prediction vectors were generated, leading to a total of 178 classifier-generated and voted prediction vectors. The optimal classification result was selected from these 178 vectors. With the use of 10-fold cross-validation, our model achieved a remarkable 97.27% overall accuracy for the 4-class classification on the study dataset. Owing to its reduced time complexity, this model is practical for CT-based PH screenings.