Narrow Results

This paper proposes a convolutional neural network for diagnosing various lung illnesses from chest CT images based on a customized Medical Image Analysis and Detection network (MIDNet18). With simplified model building, minimal complexity, easy technique, and high-performance accuracy, the MIDNet-18 CNN architecture classifies binary and multiclass medical images. Fourteen convolutional layers, 7 pooling layers, 4 dense layers, and 1 classification layer comprise the MIDNet-18 architecture. The medical image classification process involves training, validating, and testing the MIDNet-18 model. In the Lung CT image binary class dataset, 2214 images as training set, 1800 images as validation set, and 831 as test set are considered for classifying COVID images and normal lung images. In the multiclass dataset, 6720 images as training sets belonging to 3 classes, 3360 images as validation sets and 601 images as test sets are considered for classifying COVID, cancer images and normal images. Independent sample size calculated for binary classification is 26 samples for each group. Similarly, 10 sample sizes are calculated for multiclass dataset classification keeping GPower at 80%. To validate the performance of the MIDNet18 CNN architecture, the medical images of two different datasets are compared with existing models like LeNet-5, VGG-16, VGG-19, ResNet-50. In multiclass classification, the MIDNet-18 architecture gives better training accuracy and test accuracy, while the LeNet5 model obtained 92.6% and 95.9%, respectively. Similarly, VGG-16 is 89.3% and 77.2% respectively; VGG-19 is 85.8% and 85.4%, respectively; ResNet50 is 90.6% and 99%, respectively. For binary classification, the MIDNet18 architecture gives better training accuracy and test accuracy, while the LeNet-5 model has obtained 52.3% and 54.3%, respectively. Similarly, VGG 16 is 50.5% and 45.6%, respectively; VGG-19 is 50.6% and 45.6%, respectively; ResNet-50 is 96.1% and 98.4%, respectively. The classified images are further predicted using detectron-2 model and the results identify abnormalities (cancer, COVID-19) with 99% accuracy. The MIDNET18 is significantly more accurate than LeNet5, VGG19, VGG16 algorithms and is marginally better than the RESNET50 algorithm for the given lung binary dataset (Bonferroni — one-way Anova and pairwise comparison of MIDNET, LeNet5, VGG19, VGG16, and RESNET 50 ($p>0.05$)). The proposed MIDNet18 model is significantly more accurate than LeNet5, VGG19, VGG16, ResNet50 algorithms in classifying the diseases for the given multiclass lung dataset (Bonferroni — one-way Anova and pairwise comparison of MIDNET18, LeNet5, VGG19, VGG16, ResNet50 ($p>0.05$)).

Early classification and diagnosis of lung diseases is essential to increase the best chance of patient recovery and survival. Using deep learning to make it possible, the key is how to improve the robustness of the deep learning model and the accuracy of lung image classification. In order to classify the five lung diseases, we used transfer learning to improve and fine-tune the fully connected layer of VGG16, and improve the cross entropy loss function, combined with the gradient boosting decision tree (GBDT), to establish a deep learning model called a classifier. The model was trained using the ChestX-ray14 dataset. On the test set, the classification accuracy of our model for the five lung diseases was 82.43%, 95.37%, 82.11%, 79.81%, 78.13%, which is better than the best published results. The F1 value is 0.456 (95% CI 0.415, 0.496). The robustness of the model exceeds CheXNet and average performance of doctors. This study clarified that the model has strong robustness and effectiveness in classifying five lung diseases.