Deep Transfer Learning with Enhanced Feature Fusion for Detection of Abnormalities in X-ray Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Medical image classification is hindered by scarce labelled training data, impacting deep learning (DL) model performance.
• Transfer learning (TL) mitigates data scarcity but often faces domain mismatch issues, particularly with ImageNet-based models.
• Accurate classification of musculoskeletal radiographs, such as humerus and wrist X-rays, remains challenging due to data limitations.

Purpose of the Study:

• To develop a novel TL approach for medical image classification that overcomes the domain limitations of ImageNet.
• To improve the accuracy and generalisation of DL models in medical imaging tasks with limited labelled data.
• To validate the proposed TL method's effectiveness and robustness compared to conventional TL techniques.

Main Methods:

• A new TL strategy was implemented, pre-training DL models on diverse medical images similar to the target domain.
• Models were fine-tuned on small annotated medical datasets, followed by feature extraction and fusion for machine learning (ML) classifiers.
• The approach was evaluated on the MURA dataset for humerus and wrist X-ray classification, using visualization tools like Grad-CAM and LIME for validation.

Main Results:

• The proposed TL approach achieved high accuracy for humerus (87.85%) and wrist (85.58%) classification, with corresponding F1-scores and Cohen's Kappa coefficients.
• Models trained with the novel TL method significantly outperformed those using ImageNet-based TL.
• Visualization techniques confirmed the superior accuracy and robustness of the proposed method.

Conclusions:

• The developed TL approach effectively addresses the challenge of limited labelled data in medical image classification.
• The proposed method demonstrates superior performance and robustness compared to ImageNet-based TL, offering a more suitable alternative for medical applications.
• The TL and feature-fusion technique show promise for broad applicability across various medical imaging tasks, including CT scans.