Deep orthogonal multi-wavelength fusion for tomogram-free diagnosis in diffuse optical imaging

Area of Science:

• Biomedical optics
• Medical imaging
• Cancer diagnostics

Background:

• Portable diffuse optical tomography (DOT) offers non-invasive breast cancer screening.
• Distinguishing malignant from benign breast lesions is a critical challenge.
• Multi-wavelength DOT can improve signal penetration but faces challenges with correlated data.

Purpose of the Study:

• To develop an orthogonal fusion loss to regularize multi-wavelength DOT.
• To improve the accuracy of discriminating malignant from benign breast lesions.
• To create a computationally efficient model for real-time analysis on portable devices.

Main Methods:

• Implemented an orthogonal fusion loss for multi-wavelength DOT.
• Developed a raw-to-task model bypassing traditional image reconstruction.
• Validated the model on synthetic and clinical breast lesion datasets.

Main Results:

• Orthogonal fusion loss improved DOT reconstruction and lesion discrimination accuracy.
• The raw-to-task model achieved balanced accuracies of 77% (synthetic) and 66% (clinical).
• The model significantly reduced computational complexity, enabling real-time application.

Conclusions:

• Orthogonal fusion loss is effective for multi-wavelength DOT in breast cancer screening.
• Image reconstruction is not essential for accurate lesion classification using this method.
• The developed model is suitable for power-constrained, real-time portable medical devices.