Pulmonary Disease Detection from Chest x-ray Images using Dragonfly Enhanced Artificial bee Colony Hybrid Optimisation Technique

Automated analysis of chest radiographs using deep learning has shown promise for detecting pulmonary abnormalities. Aim of the current study is to develop and retrospectively evaluate a hybrid diagnostic pipeline that fuses hand-crafted and deep features, using a Dragonfly-Enhanced Artificial Bee Colony (DEABC) metaheuristic for feature selection, and classifies images with a CNN→Bi-LSTM architecture, and to compare its performance.
We used a publicly available chest X-ray collection from Kaggle containing normal and abnormal cases (pneumonia, tuberculosis, lung opacities). Images were preprocessed and lung regions segmented; texture and shape features were extracted and combined with deep CNN features. Feature selection was performed with a DEABC wrapper; classification used a CNN followed by a Bi-LSTM. We evaluated five methods (ABC, DOA, CNN, Bi-LSTM, DEABC-CNN-BiLSTM) under two hyperparameter configurations (labelled LR=0 and LR=1) using patient-level train/validation/test splits. Primary outcomes on the held-out test set included accuracy, precision (PPV), sensitivity (recall), specificity, F1 and Matthews correlation coefficient (MCC). Reported values are point estimates computed on the retrospective test split.
Across both hyperparameter configurations the DEABC-optimized CNN–BiLSTM achieved the highest point estimates on all primary metrics. For LR=0 (train ≈80%) DEABC attained accuracy 0.9552, precision 0.99597, sensitivity 0.9597, specificity 0.9805, F1 0.9703, and MCC 0.9252. For LR=1 (train ≈70%) DEABC attained accuracy 0.9504, precision 0.99099, sensitivity 0.9549, specificity 0.9756, F1 0.9655, and MCC 0.9205. Comparator models (ABC, DOA, plain CNN, Bi-LSTM) showed lower point estimates across metrics.
The proposed DEABC-enhanced hybrid pipeline produced superior retrospective discrimination on the available chest X-ray dataset relative to the evaluated baselines. These results support the potential value of combining hand-crafted descriptors, deep representations and wrapper-based feature selection—particularly when data are limited.