Variations on Branding with Text Occurrence for Optimized Body Parts Classification

Area of Science:

• Medical imaging
• Computer vision
• Machine learning

Background:

• Increasing volume of digital medical scans necessitates automated interpretation.
• Manual annotation of radiology images is time-consuming and error-prone.
• Accurate body part classification is crucial for computer-aided diagnosis.

Purpose of the Study:

• To develop an enriched multi-modal image representation for body part classification.
• To integrate automatically generated image keywords with radiographs.
• To enable combined learning by incorporating metadata into images.

Main Methods:

• Utilized Long Short-Term Memory (LSTM) based Recurrent Neural Network (RNN) Show-and-Tell model for keyword generation.
• Derived word embeddings from keywords using Word2Vec.
• Augmented radiographs with textual features via intensity markers for multi-modality.
• Trained deep learning systems with augmented radiographs.

Main Results:

• Achieved 95.78% prediction accuracy on the Musculoskeletal Radiographs (MURA) dataset.
• Obtained 83.90% prediction accuracy on the ImageCLEF 2015 Medical Clustering Task dataset.
• Demonstrated superior performance compared to existing methods.

Conclusions:

• The proposed multi-modal approach effectively enhances body part classification in radiology images.
• Combining textual features with image data offers a promising direction for medical image analysis.
• This method provides a foundation for more accurate and efficient computer-aided interpretation systems.