Enhancement of hybrid-deep learning model for microscopic bacterial image classification / Son Ali Akbar

Faculty of Electrical & Electronics Engineering Technology

Thesis (Doctor of Philosophy) -- Universiti Malaysia Pahang – 2024

Includes bibliographical references

Bacterial infection is a health problem caused by bacteria attacking all body organs (i.e., Burkholderia pseudomallei sp., Hemophilus influenzae sp., Klebsiella pneumoniae sp., Pseudomonas aeruginosa sp., Staphylococcus aureus sp., Streptococcus pneumoniae sp.). Early identification of bacterial infection significantly influences clinical microbiology diagnosis performance and decision treatment. In addition, the ability of visual microbiologists influences the traditional method based on the gold standard used to classify bacterial species through microscopy. Computer-aided diagnosis (CAD) based on image processing and machine learning is one of the modern technologies that would help clinical microbiology assessments. The current issues on the CAD implementation include the efficiency model development, and time consumption during species classification. Traditional image processing and machine learning are ineffective work its involve handcrafted feature extraction. The existing scheme applies CNN-based deep learning to achieve high-accuracy predictions. However, existing CNN algorithms face the obstacle of avoiding overfitting and computation, simultaneously taking longer. Developing a hybrid method that combines deep learning and machine learning with the efficient design of models proposes a solution to this problem. In this approach, the CNN layers are utilized for deep feature extraction. This study has proposed the enhancement of hybrid method based on deep convolution neural networks (EHDCNN) for classification the bacterial species. The EHDCNN scheme extracted the feature on the earlier CNN layer by the skip connection technique which can be efficient in the training process. Then, the SVM-based error-correcting output codes (ECOC) model was used to classify the feature map. The experiment results show that the proposed method based on an EHDCNN model produces a high prediction score, such as overall accuracy of 99.61%, recall of 99.58%, specificity of 99.97%, precision of 99.58%, F-score of 0.9958, and FPR of 0.03%, respectively. Additionally, the computational time was achieved by a rapid processing of 05.08 minutes. The high computational hardware-based GPU can ensure the speed of deep processing. In addition, GPU-based high-computing hardware can ensure deep processing speed. Therefore, the proposed model supported by highgrade devices has the potential to be applied in CAD systems.