Adaptive transfer learning for reconstructing optical-property images of elliptical phantoms

Deep learning has greatly accelerated image reconstruction in diffuse optical imaging (DOI) by enabling fast, end-to-end approximations from boundary measurements. However, models trained on simplified circular breast phantom simulations often exhibit poor generalization to complex elliptical or irregular anatomies due to domain shift, resulting in persistent reconstruction artifacts. We propose consistency-driven adaptation (CoDA), a fully unsupervised test-time adaptation method that recalibrates the batch normalization (BN) statistics of a network pre-trained on 10,000 circular phantoms. During adaptation, all convolutional weights remain frozen while BN layers are updated through gradient-based optimization with minimal learning rates to ensure stable adaptation. The process is guided by a self-supervised consistency loss that enforces consistent predictions between the original boundary measurement and its noise-perturbed counterpart. The adapted model significantly outperforms the non-adapted baseline in lesion contrast recovery, spatial detail preservation, and overall image quality based on the structural similarity index metric (SSIM) and contrast-and-size detail (CSD). Qualitative reconstructions on heterogeneous numerical breast phantoms further confirm improved suppression of boundary artifacts.