Hybrid InceptionV3-LSTM for Asthma Detection on Chest X-Rays

Asthma is commonly undiagnosed in areas without access to lung-function tests, whereas CXR is more commonly used to diagnose conditions of the lungs. The prototype is a pipeline based on InceptionV3LSTM that produces a 1024D projection from ImageNet-initialized InceptionV3 spatial features and global average pooling. A sigmoid classifier and LSTM (64) are applied to this, which has been resized into a short sequence of 1024D projections. This model is trained using a two-stage procedure that includes the previously mentioned selective fine-tuning with early stopping and frozen transfer learning in the first stage, in addition to conventional preprocessing and conservative augmentation. The dataset, comprising 4295 chest X-ray images, achieved a successful classification accuracy of 87.12%, precision of 85.64%, recall of 81.73%, F1 of 83.64%, and ROC-AUC of 0.94. This suggests that if we only concentrate on optimizing our thresholds, we can obtain higher specificity. The straightforward DenseNet-121 and ResNet-50 deep learning models can be used to illustrate the use of the sequence head. This is followed by external validation, probability calibration, and saliency audits. Referral and decision support systems are also applicable.

Hali tarjima qilinmagan