Breast tumor detection using regularized deep-learning diffuse optical tomography

Abstract

Deep-learning Diffuse Optical Tomography (DL-DOT) is a non-invasive diagnostic method that uses near-infrared radiation and deep-learning algorithms to image soft tissues in the body, such as the breast. However, DL-DOT studies have limitations, such as using only homogeneous or semihomogeneous datasets for the forward problem, which can lead to predictions not being accurate when used on experimental measurements. Another limitation regarding DL-DOT is the severe overfitting of the prediction model observed when DL methods are employed for DOT image reconstruction. To overcome this challenge, a regularized nested UNet++ deep-learning algorithm is employed. The proposed method effectively solves the DOT inverse problem in inhomogeneous breasts by applying a regularization technique. This technique reduces overfitting and simplifies the prediction model. Results show that when the regularized neural network is used to detect tumors, a minimal mean square error (MSE) loss of 5.16 × 10⁻³ is achieved compared to a non-regularized MSE loss of 4.18 × 10⁻². The enhancement of close to one order of magnitude shown by the proposed method demonstrates the significance of regularization neural networks in breast tumor detection and improving the accuracy of DOT image reconstruction.