INSTRAS: INfrared Spectroscopic imaging-based TRAnsformers for medical image Segmentation

Area of Science:

• Medical imaging
• Artificial intelligence
• Spectroscopy

Background:

• Infrared (IR) spectroscopic imaging offers rich chemical and spatial data for medical applications.
• Current machine learning models, like U-Net, struggle with the high dimensionality and long-range dependencies in IR data.
• Convolutional neural networks' inherent locality limits their effectiveness in encoding complex IR spectroscopic information.

Purpose of the Study:

• To introduce a novel deep learning model, INSTRAS, for enhanced medical image segmentation using IR spectroscopic imaging.
• To address the limitations of convolutional neural networks in capturing long-range dependencies within complex IR data.
• To evaluate INSTRAS's performance against existing convolutional models for breast IR image segmentation.

Main Methods:

• Development of INSTRAS, a transformer-based model incorporating skip-connections and transformer encoders.
• Training and evaluation of INSTRAS and various convolutional encoder-decoder models on a breast IR image dataset.
• Utilizing 9 spectral bands for segmentation tasks.

Main Results:

• INSTRAS achieved a segmentation score of 0.9788 on the breast IR image dataset.
• The transformer-based INSTRAS model demonstrated superior performance compared to purely convolutional models.
• The model effectively captures long-range dependencies, overcoming limitations of traditional CNNs.

Conclusions:

• INSTRAS represents an advanced and improved approach for segmenting medical images acquired through IR spectroscopic imaging.
• The model's ability to leverage transformer encoders enhances the analysis of high-dimensionality IR data.
• INSTRAS shows significant potential for improving diagnostic capabilities in medical imaging applications.